Changelog
What changed in each AEAntivirus release, newest first. The server, Windows client, Linux daemon and Android app each carry their own version — the notes say which components moved.
Version 2.5.257 latest
Threat Guide daily update — 1 genuinely new entry, 182 → 183 entries. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- Hashrate hijacking (mining fleet hijack) (Medium, Money-driven threats group, immediately after Cryptominer; full Spanish mirror with the same slug, order, and severity) — Luxor Technology confirmed on November 15, 2026 (CoinDesk, November 16, 2026) that an intruder used Luxor Commander, its fleet-management software, to redirect mining machines from just over a dozen customer accounts to the OCEAN pool starting around 6 p.m. Pacific on November 14. Luxor patched Commander within two hours and will pay every affected client in full; an address linked to the attack later received about 0.243 BTC (roughly $15,700). The roughly 64,000-machine / 13 EH/s figures are early operator reports Luxor has not confirmed, and the entry point is unexplained beyond an “AI-enabled vector.” Distinct from a cryptominer (malware mining on your own PC — here nothing is installed and the theft is a pool setting) and from a wallet drainer (the coins are stolen before they are earned). The entry is honest about scope: AEAntivirus does not run on mining rigs or in the fleet platform and cannot stop a pool change inside a compromised management account; it claims only the operator-workstation protections that guard the fleet login, plus the account defenses (unique password + MFA, change alerts, regular pool/address checks).
- Checked and NOT duplicated (2026-11-16 round): the FTC's November 12 fake-FTC-website alert (government impersonation + typosquatting), the trojanized Claude Code extension in the Cursor marketplace (supply-chain + malicious extension + backdoor), the Microsoft 365 outage phishing (adjudicated in 2.5.254), the Trigram campaign (ClickFix + clipboard hijacker + infostealer + job scams), the fake GTA VI PC download (game cheats + cracked software + infostealer), the Coinone vendor breach (supply-chain + credential stealer), the November 13 CISA KEV additions (exploit), and the PHP2Shell wiper chain (exploit + wiper).
- Docs in the same push: homepage “New in” rotates to 2.5.257 / 2.5.254 / 2.5.251 visible with 2.5.249 heading the fold, the Spanish homepage gains the 2.5.257 paragraph, and the admin guide gains the 2.5.257 card. The guide's “Last reviewed” stamp already reads November 2026. Severity totals: Critical 12 / High 69 / Medium 88 / Low 14.
Version 2.5.256 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 98 → 103 names. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.254.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: MongoDB, Inc. (mongodb-compass-1.53.2-win32-x64.exe, github.com/mongodb-js/compass release v1.53.2; cert SHA-1 A5BBE2A6DA1D2A6E057EF870267E6A91E4D56BAA), TeamSpeak Systems GmbH (TeamSpeak3-Client-win64-3.6.2.exe, files.teamspeak-services.com release 3.6.2 — the download link served by teamspeak.com; cert SHA-1 AD9CAF55B2599A3344BF7F9098ED8FBAAA330982), K Desktop Environment e.V. (krita-x64-5.3.4.1-setup.exe, download.kde.org stable Krita 5.3.4.1 — the download link served by krita.org; cert SHA-1 725E0BEF48AF51B490CC4BB92B2352DE5FB75475), SUSE LLC (Rancher.Desktop.Setup.1.25.0.msi, github.com/rancher-sandbox/rancher-desktop release v1.25.0; cert SHA-1 14E89EBEAF0D39AE95A0B8261CF7DEF3B4FE9250), and Florian BAUER (FreeFileSync_14.13_Windows_Setup.exe, freefilesync.org/download; cert SHA-1 84EBC099CFAB86338FEF12F19B2DC161B61B7B70).
- The spellings are the certificates', not ours: "MongoDB, Inc." carries its comma / trailing period as certified. "K Desktop Environment e.V." is Krita's signing entity — the KDE association's legal name, not the product name — with its periods exactly as certified. "SUSE LLC" is Rancher Desktop's signing entity and carries no punctuation, as certified; "TeamSpeak Systems GmbH" likewise. "Florian BAUER" is FreeFileSync's individual author signer, surname in capitals as certified — the same class as Simon Tatham (PuTTY) and Kovid Goyal (Calibre), both already on the list. Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25256) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - No verified names remain banked for the next daily run. Attempted and not added: KeePassXC (its current Windows MSI resolves to DroidMonkey Apps, LLC, already on the list), Flameshot (resolves to the shared SignPath Foundation signing service — excluded on scope grounds, the 2.5.242 Transmission precedent), and CopyQ (its official installer is unsigned).
- Docs in the same push: admin guide gains the 2.5.256 card above the 2.5.255 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.254).
Version 2.5.255 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 93 → 98 names. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.254.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Takashi Sawanaka (WinMerge-2.16.60-x64-Setup.exe, github.com/WinMerge/winmerge release v2.16.60; cert SHA-1 4F8C8F56C77DCC6C96FE293589E06E4411C230C9), draw.io Ltd (draw.io-32.6.10-windows-installer.exe, github.com/jgraph/drawio-desktop release v32.6.10; cert SHA-1 145B546AD357C6D87CF6F702D7BFCB2876635C47), Logseq, Inc. (Logseq-win-x64-2.0.2-nsis.exe, github.com/logseq/logseq release 2.0.2; cert SHA-1 D1799FCD7EB901EA03D378EDAE5CC4C4CF21CAEA), Red Hat, LLC (podman-desktop-1.30.5-setup-x64.exe, github.com/podman-desktop/podman-desktop release v1.30.5; cert SHA-1 9F2794A5AAD1DC9F977B640AEF10745C208E72EC), and Bruno Software Inc (bruno_4.3.0_x64_win.exe, github.com/usebruno/bruno release v4.3.0; cert SHA-1 B83EF0282C0D11C66C8A611E9606A1A3BAD9DD4A).
- The spellings are the certificates', not ours: "Logseq, Inc." carries its comma / trailing period as certified. "Red Hat, LLC" is Podman Desktop's signing entity — the certificate's legal name ships, never the product name — and carries its comma with no period after LLC. "Bruno Software Inc" and "draw.io Ltd" carry no suffix punctuation, as certified. "Takashi Sawanaka" is WinMerge's individual author signer — the same class as Simon Tatham (PuTTY) and Kovid Goyal (Calibre), both already on the list. Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25255) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - No verified names remain banked for the next daily run.
- Docs in the same push: admin guide gains the 2.5.255 card above the 2.5.253 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.254).
Version 2.5.254
The Threat Guide's daily update — 1 genuinely new entry, 181 → 182, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): Autonomous AI attacks (AI-run hacking) (High, Stealth & network threats — FBI FLASH-20261113-001, November 13, 2026): attackers run open-source AI penetration-testing frameworks that take a goal in plain language and hand the work to a crew of AI agents — mapping the target, finding a way in, climbing to administrator rights, moving sideways, and copying data out. The FBI says the framework ARTEX has been used this way since at least August 2026 against financial services, government, defense, technology, semiconductors, energy, healthcare, manufacturing, telecommunications, and law, and that some intrusions went from first targeting to privileged access in under three hours. ARTEX ships no model of its own — the operator plugs in Claude, GPT, DeepSeek, GLM, Grok, or Kimi (attacker copies mostly use Chinese models, led by DeepSeek Flash); its developer released it in July 2026 for authorized testing and removed it on October 8 after abuse reports, but forks and a paid edition circulate, and more than 1,600 control consoles were exposed online in early November. Cited campaigns include South Korean financial firms (personal data on about 67,000 people) and a ConnectWise Cloud Backup compromise reaching more than 2,250 downstream tenants; a second framework, Cairn, was reported the same week in China-linked use. Distinct from AI agent instruction hijacking (a planted file tricks YOUR agent — here the agents are the attacker's and there may be no lure or click at all) and from malicious AI models (a downloaded model is the trap — here the AI is the burglar, not the bait).
- Honest scope: the entry says plainly that no antivirus can slow an attacker's AI and that software on a PC cannot defend a supplier's server it does not run on. It claims only the endpoint line AEAntivirus actually ships — signature checks on the tools and payloads the agents run, behavioral and API-sequence scoring of exploitation, privilege-climbing, and data-copying behavior, Exploit Protection, the software updater closing the unpatched flaws the agents hunt first, and network monitoring plus exfiltration-anomaly detection — and gives the FBI-shaped defenses for a three-hour break-in: patch forgotten and internet-facing services first, multi-factor authentication on every remote and partner login, least-privilege accounts, never expose a testing or admin console to the internet, and be ready to respond in hours. Severity totals are now Critical 12 / High 69 / Medium 87 / Low 14.
- Checked and NOT duplicated: the Microsoft 365 outage phishing (Microsoft Threat Intelligence, November 13–14, 2026 — fake service-health emails citing the real incident EX1550543 and a "RESTORE MESSAGES" button; a current-event lure on the CrowdStrike-outage precedent — stays inside phishing + tech-support scams). The Trezor Safe 5 implant confirmations (Trezor, November 12, 2026 — the same CryptoBilis case the tampered hardware wallet entry already carries by name). The Macfinger / DownTheDRAIN macOS infostealer (SANS / Huntress, November 13, 2026 — fake-CAPTCHA paste-into-Terminal; stays inside ClickFix + infostealer + wallet stealer). TypoPrompting lookalike AI-tool domains (Island, November 2026; stays inside typosquatting + SEO poisoning + ClickFix). The AI-triaging infostealer that sends stolen data to Claude for ranking (November 2, 2026 — post-theft sorting that changes neither entry, detection, nor advice; stays inside infostealer + AI assistant credential stealer). The Bermuda Google Meet bank calls (Bermuda Police Service, November 13, 2026; stays inside vishing + calendar invite phishing + fake meeting invites). The Escambia County Amendment 3 and Frederick County property-tax warnings (November 10–13, 2026; stay inside government impersonation, on the Minnesota property-tax precedent of 2.5.249). The refrigeration-controller joint advisory (CISA / FBI, November 5, 2026 — the EU supermarket campaign already adjudicated on 2026-11-03; stays inside router & IoT malware + exploit). The Check Point PHP2Shell ESXi wiper, the N-central exploitation, and the November 13 CISA KEV additions (stay inside exploit / zero-day attack + wiper / malicious remote tools). The Neutrino ransomware family (a new name on double-extortion behavior; stays inside ransomware + double extortion). And the fake security tools on GitHub (Dr.Web via BleepingComputer, November 6, 2026; stay inside supply-chain attack + malicious packages).
- Counts stay live: guideCount is computed from the guide data (182), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.254 above 2.5.251 / 2.5.249 (2.5.248 folds in, and the fold summary now names Server 2.5.248); the ES homepage gains the 2.5.254 paragraph; the admin guide gains the 2.5.254 card above the 2.5.251 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.253 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 88 → 93 names. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.251.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Kandra Labs, Inc. (Zulip-5.14.2-x64.msi, github.com/zulip/zulip-desktop release v5.14.2; cert SHA-1 9221BAADA4B49618E4890C83D3A3AE9636F6FBF4), Eugen Pankov Softwareentwicklung (tabby-1.0.238-setup-x64.exe, github.com/Eugeny/tabby release v1.0.238; cert SHA-1 A04699ED7E01AFBDB99AE9E93204956801BE6B3C), Skymatic GmbH (Cryptomator-1.19.6-x64.msi, github.com/cryptomator/cryptomator release 1.19.6; cert SHA-1 C0F1E9D2FEAED456C8DB138E1753CD7ADD1CB484), Meltytech, LLC (shotcut-win64-26.9.27.exe, github.com/mltframework/shotcut release v26.9.27; cert SHA-1 796DC3140F9A322B07BCA512A86EA4E37B79AB25), and Joplin (Joplin-Setup-3.7.22.exe, github.com/laurent22/joplin release v3.7.22; cert SHA-1 AF0C1B9451567F9AF077DD0646D2927153B05DAF).
- The spellings are the certificates', not ours: "Kandra Labs, Inc." and "Meltytech, LLC" carry their comma / trailing period as certified. "Eugen Pankov Softwareentwicklung" is Tabby terminal's full German signing-entity name as certified, with no abbreviation. "Skymatic GmbH" is Cryptomator's signing entity — the certificate's legal name ships, never the product name. "Joplin" is the single word its certificate carries as both common name and organization — the same product-named class as VideoLAN and NOTEPAD++ already on the list. Kandra Labs, Inc. and Eugen Pankov Softwareentwicklung were banked by the 2.5.250 run's 5-per-day cap and are shipped today only after re-downloads that were byte-identical to the banked artifacts (same size, same SHA-256, same cert SHA-1). Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25253) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - No verified names remain banked for the next daily run. Attempted and not added: Standard Notes (standard-notes-3.202.11-win-x64.exe, github.com/standardnotes/app) — its current Windows installer resolves to Proton AG, which is already on the list, so there is nothing new to add.
- Docs in the same push: admin guide gains the 2.5.253 card above the 2.5.250 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.251).
Version 2.5.252 (admin-only)
Installer and update downloads move off the server and onto GitHub. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.251.)
- Why: on November 12 the Render workspace used its full 5 GB Hobby-plan bandwidth allowance (overage $0.15/GB until December 1), and the driver was this service: the installer/update binaries are the only multi-megabyte files it serves — about 15.4 MB per full set (both Windows installers, both delta patches, the Android APK, the Linux and macOS tarballs) — while everything else it serves is kilobytes (the signature database is 1.3 KB).
- What changed: the binaries now live in the public AMott-2020/AEAntivirus-Releases GitHub repo (binaries only — the source repo stays private). When the host env var
BINARY_CDN_BASE is set to that repo's raw base URL, the served update manifests rewrite their binary URLs (Windows exe_url / exe_x64_url, both delta patch_urls, and the Linux/macOS/Android exe_urls) to the CDN address in the response — the disk manifests are never touched — and the account-gated /download/file/* and /download/AEAntivirus-Setup.exe routes answer with a redirect to the CDN after the usual account and license checks. /updates/proxy/* stays as the fallback route. - No client change was needed: the shipped Windows client's
make_abs_url() and the Linux daemon's make_url() both pass absolute http(s) URLs through unchanged, and both downloaders follow redirects (WinHTTP default policy; libcurl FOLLOWLOCATION), so already-installed clients pick up CDN URLs from the next manifest they fetch — manifests are served no-cache, must-revalidate, so there is no stale-manifest window. The Android app never downloads its APK from the manifest (only the /download page serves it). SHA-256 verification is unchanged — the CDN files are byte-identical to the server copies, so the manifest hashes still match. - What stays on the server: the manifest JSON itself (it must — the Free-plan daily-signature-snapshot rewrite and the beta channel are response-side), the signature databases, certs, imphash, and trusted-signer lists (all kilobyte-scale), plus every page and API.
- Cutover and rollback: with
BINARY_CDN_BASE unset, behavior is exactly the pre-2.5.252 behavior. Cutover = publish the binaries (below), set the env var on Render, redeploy. Rollback = unset it and redeploy. - Pipeline:
server/scripts/publish-binaries-github.sh pushes the seven binary files from server/updates/ to the releases repo (via the stored GitHub credential's push helper) and verifies the pushed sizes against the Git tree; build.sh and publish-update.sh call it (non-fatally) at the end of every publish, so future client releases land on the CDN host automatically. One-time setup: the GitHub fine-grained token must list AMott-2020/AEAntivirus-Releases under its repository access before the first publish can succeed — the script prints exactly that instruction when the push is refused. - Docs in the same push: README gains this section; the admin guide gains the 2.5.252 card above the 2.5.251 card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.251). Client What's New is untouched.
Version 2.5.251
The Threat Guide's daily update — 1 genuinely new entry, 180 → 181, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): FaceStealer fake video-call scam (replayed webcam calls) (High, Tricks, traps & nuisances — FBI FLASH AC-111226-MS, November 12, 2026; Justice Department disruption the same day; Reuters and Help Net Security, November 12–13, 2026): North Korean actors tracked as Sapphire Sleet, UNC1069, and APT38 stole more than $400 million in cryptocurrency over 18 months by luring cryptocurrency and Web3 workers and open-source maintainers into imposter Zoom, Teams, and Meet calls populated with secretly recorded webcam clips of earlier victims — the FBI seized more than 900 such videos, a library replayed to each victim's own contacts. The invite comes from the victim's real hijacked Telegram, LinkedIn, Slack, or Discord account, tailored from their chat history; a Calendly booking follows, the meeting link arrives shortly before the call on a lookalike domain, the target's webcam records from the click, and a fake "cannot hear you" audio fault walks the target through an update, a script, or a pasted terminal command that installs a cookie harvester, keyloggers, a screen capturer, and a backdoor — with theft sometimes delayed from a few months to more than a year, until wallet balances grow. Distinct from fake meeting invites (a stranger's invitation leads to a fake meeting app you download — here the invitation comes from your contact's real, hijacked account, the face is that contact's real recorded image, and the payload is a fake audio fix delivered inside a live-looking call) and from deepfake scams (an AI clone pretends to be someone — here no clone is needed, because the recording is the person; the FBI says genuine stolen footage still overwhelmingly dominates the scheme, with deepfakes of well-known crypto figures only beginning).
- Honest scope: the entry claims only the payload and link layers AEAntivirus actually ships — signature checks on the installed files, behavioral and API-sequence scoring of the cookie harvesting, keylogging, screen capture, and backdoor behavior, and the phishing layers and DNS feeds against known imposter meeting domains — and says plainly that no software can tell whether the familiar face on a video call is live or a recording, because a replayed clip of a real friend passes every camera check. The defenses given are the FBI's: join through the official Zoom or Teams app with an invite code, not a meeting link; keep your webcam off until you can see, hear, and talk with the person who invited you; confirm an unexpected meeting through a second channel; never run a script, paste a command, or install an "update" to fix a meeting; keep only small amounts in browser-extension wallets and never store a seed phrase or private key on a device; and if you think you were caught, move crypto at once to a new wallet created on a clean device, change passwords, log out of every session, warn your contacts, and report to IC3 with the keyword "FaceStealer" (fbi.gov/facestealer). Severity totals are now Critical 12 / High 68 / Medium 87 / Low 14.
- Checked and NOT duplicated: wfx-loot, the GitHub CI secret-theft campaign (Aikido via BleepingComputer, November 12, 2026 — 4,837 compromised accounts probing 61,764 repositories, pull requests titled "docs: dependency audit notes," the wfx-docs branch; a campaign name whose mechanism was already adjudicated as GhostAction in 2.5.248 — it stays inside supply-chain attack + credential stealer, and BleepingComputer itself notes the overlaps with GhostAction). SCKit, the ScreenConnect worm (already adjudicated in 2.5.249; not re-litigated). XLoader 8.9 delivered by purchase-order DOCX (AhnLab ASEC, November 12, 2026 — a malware family name riding an old Equation Editor chain; stays inside infostealer + exploit + macro malware). The fake internet-provider speed-test call that installs UltraViewer (Toronto Police, November 6, 2026; stays inside remote-access scams + tech-support scams). The City of Brandon, South Dakota invoice-by-wire warning (November 13, 2026; stays inside the zoning & permit fee scam + government impersonation). The fake job interview that registers a business in the victim's name (a single unverified account, November 12, 2026; stays inside job scams). The November 9–10 CISA KEV additions (stay inside exploit / zero-day attack). And the separate TraderTraitor Binance matter in the same November 12 Justice Department release (already adjudicated; stays inside job scams + fake meeting invites).
- Counts stay live: guideCount is computed from the guide data (181), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.251 above 2.5.249 / 2.5.248 (2.5.247 folds in, and the fold summary now names Server 2.5.247); the ES homepage gains the 2.5.251 paragraph; the admin guide gains the 2.5.251 card above the 2.5.249 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.250 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 83 → 88 names. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.249.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Proton AG (ProtonVPN_v5.1.9_x64.exe, github.com/ProtonVPN/win-app release v5.1.9; cert SHA-1 F0B999E273EAFDABCDA5A3B86315E5604DFB9A49), Mattermost, Inc. (mattermost-desktop-6.3.2-win-x64.msi, github.com/mattermost/desktop release v6.3.2; cert SHA-1 2F7A2C23057919D2DE2B03DDB0ADB6C64EB409B7), Anki Software LLC (anki-26.09.4-win-x64.msi, github.com/ankitects/anki release 26.09.4; cert SHA-1 CD0E64BECCCCF95B15F70D0B4174F6452FE86161), Tien Do Nam (LocalSend-1.18.2-windows-x86-64.exe, github.com/localsend/localsend release v1.18.2; cert SHA-1 D95D73B75580AF7B614F33F68D0EB1B5EBD65600), and Rocket.Chat Technologies Corp. (rocketchat-4.19.3-win-x64.exe, github.com/RocketChat/Rocket.Chat.Electron release 4.19.3; cert SHA-1 15F2A2245BE1131274E66E3C55EC8D04C89CACCF).
- The spellings are the certificates', not ours: "Proton AG" is Proton VPN's signing entity — the vendor sought since 2.5.228, whose own file host never served a binary through the build's network route; today the project's official GitHub release channel served it and the signer resolved cleanly, no suffix or punctuation as certified. "Mattermost, Inc." and "Rocket.Chat Technologies Corp." carry their comma / trailing period as certified; "Anki Software LLC" carries no punctuation. "Tien Do Nam" is LocalSend's individual author signer — the same class as Kovid Goyal (Calibre), Simon Tatham (PuTTY) and Krzysztof Kowalczyk (Sumatra PDF) already on the list: the software is widely run, and the certificate's legal name is what ships, never the product name. Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25250) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but held by the standing 5-per-day cap (banked for the next daily run): Kandra Labs, Inc. (Zulip-5.14.2-x64.msi, github.com/zulip/zulip-desktop release v5.14.2; cert SHA-1 9221BAADA4B49618E4890C83D3A3AE9636F6FBF4) and Eugen Pankov Softwareentwicklung (tabby-1.0.238-setup-x64.exe, github.com/eugeny/tabby release v1.0.238; cert SHA-1 A04699ED7E01AFBDB99AE9E93204956801BE6B3C). Attempted and not added: Jellyfin Media Player (its official Windows installer is unsigned — the same settled class as 7-Zip, Kodi and ShareX); Nextcloud desktop and IVPN desktop (their GitHub latest releases carry no Windows installer assets, and no artifact was obtained from an official channel this run); WinSCP and Shotcut (no GitHub latest release resolves for either project).
- Docs in the same push: admin guide gains the 2.5.250 card above the 2.5.242 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.249).
Version 2.5.249
The Threat Guide's daily update — 3 genuinely new entries, 177 → 180, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 3 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Zoning & permit fee scam (fake planning officials) (Medium, Tricks, traps & nuisances — FBI PSA260309, March 9, 2026, tracked since December 2025, with county warnings still landing in November 2026, Horry County, South Carolina on November 3 among them): criminals mine publicly available permit information — real case numbers, property addresses, applicant names, and the true names of planning officials — and email people with an active land-use, zoning, or building permit application an official-looking invoice, often a PDF, for an "approval fee" payable at once by wire, payment app, or crypto, from non-government domains such as @usa.com. Distinct from government impersonation generally (the threat there is invented — here the application, the case, and the fee are real; only the collector is fake). Memo of sale scam (cash-buyer title cloud) (Medium, Tricks, traps & nuisances — joint Tennessee alert, November 10, 2026: TDCI, the Tennessee Real Estate Commission, the Attorney General's Office, Tennessee REALTORS, ALTA, the Tennessee Bankers Association, and the County Officials Association): a "we buy houses for cash" company gets a contract signed, records a memorandum of sale on the home's title instead of closing, then refuses to close and demands typically tens of thousands of dollars to remove the memo; some operators are wholesalers who never meant to own the home. Distinct from home title (deed) fraud (a stranger forges a deed — here nothing is forged, the owner signed willingly, and the weapon is a legitimate filing the owner never understood they had authorized). And Screen-share scam (WhatsApp screen mirroring) (Medium, Tricks, traps & nuisances — Which?, September 10, 2026, republished November 12, 2026): a caller posing as a bank, the police, or the FCA moves the victim to a WhatsApp video call and into screen sharing, then simply watches — banking apps, typed passwords, and live one-time codes — while the victim makes every tap; one man was kept on calls for 20 hours over three days and lost more than £27,000. Distinct from the remote-access scam (a tool is installed and the stranger drives — here nothing is installed and no control is handed over, because watching is enough).
- Honest scope: the zoning entry claims only the message layers (email scanning grades the pattern; web protection blocks known scam payment pages) and says plainly that no software can tell whether a fee is really owed on a real application — the defenses are the FBI's (.gov sender domains, no permit fees by wire through email, call the planning department yourself on its published number, pay only through the official portal or in person, call the bank at once and report to IC3 if you paid). The memo-of-sale entry claims no software counter at all — nothing touches a computer and the damaging act is a paper filing under a signed contract — and gives the Tennessee warning's defenses (read for the word "memorandum," get removal terms and any wholesale plan in writing, check the buyer's license, have a real-estate attorney review before signing, use the county recorder's free property fraud alert, and never pay to release a memo). The screen-share entry claims only the upstream phone layers (Scam-call Blocker, Senior Protection Pack warnings) and says plainly that no software can stop it once the call is running — WhatsApp and screen sharing are both legitimate — so the defenses are absolute rules (no bank, regulator, or police officer will ever ask you to share your screen; never share with banking, messages, or email open, because a watcher sees one-time codes live; hang up and call your bank on the number on your card; if you already shared, stop, call the bank, and change seen passwords from another device). Severity totals are now Critical 12 / High 67 / Medium 87 / Low 14.
- Checked and NOT duplicated: SCKit, the ScreenConnect worm (Huntress, November 11, 2026 — a toolkit / campaign name; its mechanisms stay inside worm + abused remote-access tools + session hijacking, on the GhostAction / Mirax precedent, and its technician-trap takeover was designed but not observed to succeed). The genuine Google security email carrying a fake $399 McAfee bill (University of Michigan, October 6, 2026; reports November 2 and 9 — a genuine sender channel carrying attacker text whose payload is a phone number; stays inside real order-email scams + callback phishing, on the Trezor precedent of 2.5.210). The CISA / FBI Russian messaging-app phishing recirculation (November 11–12, 2026; the advisory family dates to March / June 2026 — "Signal Support" impersonation, QR / link device linking, and PIN / code surrender stay inside QR login jacking + verification-code forwarding). The FTC Prime settlement-fund messages (November 10, 2026 — already named inside the settlement & class-action refund scams entry itself). The FTC health-insurance open-enrollment alert (October 29, 2026 — misleading "comprehensive" plans and discount cards stay inside Medicare & health-benefits scams, whose entry already carries the seasonal open-enrollment cousins). The Minnesota senior property-tax exemption grant email (Wilkin County Sheriff, November 11, 2026 — a nonexistent exemption; stays inside government impersonation, on the property-tax precedent of 2.5.248). The FBI Austin-area teen sextortion warning (November 10, 2026; stays inside sextortion). The fake traffic-violation QR texts, the fake AI event pages, and the DocuSign ScreenConnect lure were already adjudicated and were not re-litigated.
- Counts stay live: guideCount is computed from the guide data (180), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.249 above 2.5.248 / 2.5.247 (2.5.246 folds in, and the fold summary now names Server 2.5.246); the ES homepage gains the 2.5.249 paragraph; the admin guide gains the 2.5.249 card above the 2.5.248 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.248
The Threat Guide's daily update — 2 genuinely new entries, 175 → 177, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 2 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Inmate-release / ankle-monitor scam (Medium, Tricks, traps & nuisances — Boone County, Missouri circuit court, November 10, 2026; Cleburne County Sheriff, Arkansas, November 10, 2026; DeSoto Parish Sheriff, Louisiana, November 11, 2026): callers mine the public jail rosters and arrest reports for real, fresh arrests, find the inmate's family and friends, and pose as jail staff, court officials, or deputies to sell an immediate release to home detention or electronic monitoring for an upfront fee for an ankle bracelet or monitoring device. The court in Boone County — which calls it the "home detention scam" and counts at least three confirmed instances — says the callers used real judges' names and actual booking details; in DeSoto Parish the caller knew a real arrest's specifics and a deputy's name and badge number and collected more than $1,000. Distinct from family emergency ("grandparent") scams (the arrest is invented and the "relative" does the begging — here the arrest is real and verifiable, which disarms the one check families usually try, and the caller is an official) and from government impersonation scams (the threat there is the victim's own arrest — here the offer is to buy someone else's release). And Barcode payment scam (pay at the register) (Medium, Tricks, traps & nuisances — Social Security Administration Office of Inspector General alert, September 28, 2026; Ohio Attorney General consumer newsletter, June 2026; Horry County Sheriff, South Carolina, and DeSoto Parish Sheriff, Louisiana, November 2026): a scary call or text — a warrant, a fine or debt, a family member's bail, a utility shutoff — and then, instead of gift-card numbers or a wire, a barcode or QR code to be taken to a big-name retailer or pharmacy and loaded with cash at the register. The scan is the payment: the money lands in the scammer's account, the store is the unwitting cash register, and there is no charge to dispute and no transfer to recall; the SSA inspector general notes victims lose thousands often without ever sharing personal information — "the barcode itself is the scam." Distinct from gift card scams (the victim buys a card and reads out its numbers — here no card exists and the store clerk is the unwitting payment processor) and from QR-code phishing / quishing (the victim scans a code with their own phone and lands on a phishing page — here the store scans the code, and the code is not a link but the payment instrument itself).
- Honest scope: the inmate-release entry claims only the upstream phone layers — the Scam-call Blocker blocking known scam numbers, and the Senior Protection Pack's scam-call warnings and trusted-contact alerts — and says plainly that AEAntivirus cannot sit in on a phone call and no software can release anyone from jail; the defenses given are the courts' own published rules (no court or jail will ever collect a release fee over the phone, and no agency takes gift cards, payment apps, prepaid cards, or crypto for one; any real home-detention fee is handled in person through the court clerk's office; a caller who knows the booking details has proved nothing, because those details are public — that is how they got them; check any release, bail, or monitoring claim by calling the jail, the clerk, or the inmate's lawyer yourself). The barcode entry says plainly that no software can intercept a cash deposit handed to a cashier; the defenses given are the SSA inspector general's rule (law enforcement will never demand payment by sending a barcode, and neither will any government agency, court, or utility — a barcode sent by a caller is not a bill, a bond, or a government payment channel) plus verifying the underlying claim on a number you look up yourself, and acting the same day if you already paid.
- Checked and NOT duplicated: The joint refrigeration advisory, GhostMap, and SafePay (already adjudicated in earlier rounds and not re-litigated). The November 9–10 CISA KEV additions and the week's vulnerability news — Barracuda ESG exploitation, SAP's ICM2Shell / PROXYGATE fixes (no active exploitation established), Microsoft's 1,093-flaw Patch Tuesday including the exploited Outlook spoofing flaw CVE-2026-124194, and the Chrome zero-day (all stay inside exploit / zero-day attack). The fake AI event pages — OpenAI, SpaceX, and Anthropic registration pages (Push Security, November 5, 2026; already adjudicated in the 2.5.243 / 2.5.245 rounds; their mechanisms stay inside device-code phishing + adversary-in-the-middle phishing + fake meeting invites). Operation CloudFire / Nyasher calendar-invite session hijacking (already adjudicated; stays inside adversary-in-the-middle phishing + calendar invite phishing + session hijacking). The fake traffic-violation text campaign with QR codes set in a court-notice image (November 10, 2026; stays inside QR-code phishing + government impersonation, on the traffic-ticket precedent of 2.5.208). The Veterans Day roundups, including the FTC's November 10, 2026 social-media reminder (fake VA grant programs built from a veteran's own profile) and California's November 10 claim-sharks alert (the roundup family was adjudicated in 2.5.246; stays inside government impersonation + charity scams, with the paid-helper pattern already carried by the student loan and Medicare entries). GhostAction, the GitHub Actions workflow campaign sweeping repository and git-history secrets (Socket / GitGuardian, November 9–10, 2026; a campaign name — stays inside supply-chain attack + credential stealer). The strutil-core malicious Ruby gem (Socket, November 6, 2026; an install-time credential worm that poisons AI coding-tool configs — stays inside malicious software packages + AI agent instruction hijacking; no confirmed victims, removed within about five hours). Mirax, the Android banking RAT spread through streaming-app ads (Group-IB, November 5, 2026; a malware family name — stays inside banking trojan + accessibility-service abuse + fake & malicious phone apps). The DocuSign lure delivering ScreenConnect remote-access software (MailGuard, November 11, 2026; stays inside abused remote-access tools + phishing + fake shared-file lures). The Across Protocol domain hijack via registrar social engineering (November 10, 2026; stays inside DNS hijacker + crypto wallet drainer, on the DeFi domain-hijack precedent of 2.5.246). The dormant-wallet address-poisoning attempt (CoinDesk, November 10, 2026; already the address poisoning entry). The Workday payroll session-hijack advisory (Fluency, November 9–10, 2026; payroll diversion stays inside spear phishing & BEC, the session persistence inside session hijacking, and Fluency's own new detection targets the MFA downgrade step, which is its own entry). The holiday fake-store surge reporting (Detroit Free Press, November 11, 2026 — some 500 new fake stores a day by one vendor's count; a surge statistic for fake online stores, not a new script). The McHenry County property-tax payment-plan calls (November 10–11, 2026; stay inside government impersonation + fake debt-collector scams). The Wisconsin gold-courier retirement loss (November 11, 2026; stays inside gold-bar courier scams). The sapho11 wallet-developer supply-chain campaign (November 10, 2026; stays inside supply-chain attack + malicious software packages, on the Keplr adjudication of 2.5.245). PhantomDOM and the fake GitHub security tools were already adjudicated and are not re-litigated here.
Version 2.5.247
The Threat Guide's daily update — 2 genuinely new entries, 173 → 175, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 2 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Settlement & class-action refund scams (Medium, Tricks, traps & nuisances — FTC Amazon refunds page and the FTC's September 17, 2026 revised-order announcement): the real Prime refunds are automatic — $1.5 billion, up to $200 per person under the revised order, more than $845 million already paid by September 2026, payments continuing through April 2027 with a new automatic round from October 1, 2026 for the 11–20-benefits group — with no claim form, no fee, and no action needed, and the FTC states plainly that it is not contacting people about these refunds and will never ask anyone to pay to get one; the scam borrows the settlement's name (the November 10, 2026 warnings name messages about an "FTC Prime subscription settlement fund") and invents the claim step the real program does not have — a form harvesting card, bank, and identity details, a "verification" asking for an account password, or a processing / release fee. Distinct from refund scams (a staged overpayment the victim is guilted into sending back), prize & lottery scams (fees to claim a prize never entered — here real money really is being paid automatically to millions, which is exactly what makes the fake claim plausible), and recovery scams (paid help recovering scam losses). And Student loan scams (forgiveness & consolidation traps) (Medium, Tricks, traps & nuisances — New York Attorney General consumer alert, November 4, 2026, building on her September warning about fraudsters using student-loan program changes): since July 1, 2026, under the One Big Beautiful Bill Act, a new consolidation loan treats the borrower as a new borrower and can permanently limit repayment options and cut off paths to forgiveness; relief companies and callers promise guaranteed or immediate forgiveness or pressure borrowers to consolidate now before a deadline, charge upfront or monthly fees for applications and forms the Department of Education and the borrower's own servicer provide free, and harvest the Federal Student Aid ID and bank details — and a consolidation signed under that pressure cannot be undone. Distinct from fake debt-collector scams (a collector demanding payment on a debt that does not exist — here a "helper" sells relief for a real debt) and from Medicare & health-benefits scams (a different benefits system, a different signature).
- Honest scope: the settlement entry claims only the layers that touch it — email scanning grading the impersonation pattern, DNS filtering blocking known fake claim sites, web protection scoring claim pages — and says plainly that no software can sit on a phone call; the defenses given are the settlement's own published rules (no action is needed to receive a real refund, so any message demanding action is lying; never pay to get a refund; check only through channels you reach yourself — the FTC refunds page typed in by hand, your own PayPal, Venmo, or mail — and check promptly, because real payments expire 60 days after issue, the very deadline the fake "expiring claim" messages imitate). The student-loan entry says plainly that no antivirus can judge a sales pitch on the phone and no software can undo a consolidation you signed; the defenses given are the free ones (everything the companies charge for is free through your servicer and studentaid.gov; never pay upfront; never share your FSA ID or password; get free, unbiased advice before consolidating — in New York, the Education Debt Consumer Assistance Program — because since July 1, 2026 the signature is permanent).
- Checked and NOT duplicated: the FTC's imposter-ads alert (November 2, 2026; stays inside malvertising + SEO poisoning — whose entry already names bought ads and fake support pages — + fake online stores). The fresh FTC phone-deactivation PSA (already its own entry since 2.5.208). MadApe, the PDF-lure device-code phishing kit (Group-IB, October 29, 2026; stays inside device-code phishing — a kit name for a covered mechanism). BluePhishing agent-blueprint consent (Proofpoint; a research proof of concept, expressly not yet seen in the wild; stays inside OAuth consent phishing). The recirculated fake-hospital / Chinese law-enforcement transfer script (the FBI's November 13, 2025 PSA I-111325-PSA; stays inside government impersonation + digital arrest + Medicare scams). The FBI Honolulu law-enforcement impersonation warning (November 2, 2026; stays inside government impersonation). The NIKKE PC launcher tampering (BleepingComputer, November 4, 2026; stays inside supply-chain attack + malicious game cheats). Turla's Cereal Exchange toolset (ESET via BleepingComputer, November 10, 2026; a state espionage toolset, stays inside spyware — not a consumer threat type). The Vanderbilt / BYOD ransomware claim (November 9, 2026; an unverified group claim; stays inside ransomware + double extortion). The CISA KEV additions of November 9, 2026 (GPUStack, FreeSWITCH, OpenSSH, Aspera) and the PowerSYSTEM Center advisory (stay inside exploit). PhantomDOM, the fake GitHub security tools, the joint refrigeration advisory, and SafePay were already adjudicated and are not re-litigated here.
Version 2.5.246
The Threat Guide's daily update — 2 genuinely new entries, 171 → 173, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 2 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Fake gift card giveaway (verification trap) (Medium, Tricks, traps & nuisances — CTM360's November 2026 Gift Card Report, via BleepingComputer, November 7, 2026): more than 1,000 incidents of one campaign design impersonating PlayStation, Walmart, Amazon, Sephora, and SHEIN, often with fabricated celebrity endorsements (MrBeast and Alix Earle among the names used); the portal lets the victim pick an amount, simulates gift-card code generation, and shows a partially masked code (XXXX-XXXX-7F3A-XXXX) as proof, then gates the final digits behind a "human verification" step that loops through surveys, quizzes, app installs, and offers — every completed step reported as incomplete — and in every flow CTM360 tested, no gift card was ever delivered; the operators earn affiliate commissions when a victim submits a mobile number and enters the PIN sent to it, a step that enrolls the victim in a paid subscription billed through their phone carrier. Distinct from gift card scams (the victim is told to pay IN gift cards), gift card draining (a real card emptied before use), social-media giveaway scams (a crypto-doubling pitch), and prize & lottery scams (fees demanded to claim a prize): here the victim pays the scammer nothing directly and no credential is the point — their time, personal details, installs, and a subscription enrollment are the product. And Restaurant booking scam (fake group booking) (Medium, Tricks, traps & nuisances — New Zealand Police, November 6–9, 2026): callers pose as customers arranging large group or function bookings at Auckland restaurants — so far mostly Chinese restaurants, ahead of the festive season — on condition the restaurant itself pre-pays a special bulk order (in the reported cases abalone from a supplier in China) to the supplier's offshore account; the "customer" then sends a fake bank transfer confirmation so it looks as though the restaurant has been reimbursed or the booking paid for, the restaurant wires the money, and neither booking nor supplier was real; one Auckland restaurant lost $15,000, and the Auckland City Financial Crime Unit said several restaurants reported the same approach in a matter of weeks. Distinct from spear phishing & BEC (a real supplier or boss impersonated — here the "customer" invents both the booking and the supplier) and from marketplace scams' fake-overpayment refunds (a consumer seller sending a difference back — here a business pays a third party at a customer's direction before any service).
- Honest scope: the giveaway entry claims only the follow-on coverage that exists — web protection and DNS filtering block the known reward portals and the message layers grade the lures — and says plainly that no filter can make a fake card real; the defenses given are the rules (a genuine card's complete code is simply given, never "verified," surveyed for, or unlocked by installing apps; a PIN sent to your phone on a reward page is what authorizes a carrier-billed subscription, so check the phone bill and cancel with the carrier). The restaurant entry claims no software counter for the core act, because there isn't one — no antivirus can judge a booking taken over the phone, and the payment is an ordinary bank transfer the restaurant chooses to make; only the written version is covered, by email scanning grading the pattern. The defenses given are the business ones (never pre-pay a supplier a customer chose; a transfer confirmation sent by the customer is not money — money is only real as cleared funds in your own bank account, checked in your own banking app or statement; call the bank the same day if money already went).
- Checked and NOT duplicated: the WhatsApp "vote for me" takeover wave (INCIBE, October 21, 2026; Swiss NCSC, October 27; Gardaí "Vote for Mia," November 4; Derbyshire Police, November 7) — its two traps are the two entries that already exist: the QR scan is QR login jacking, whose entry already names hijacked friends' accounts spreading the same trap, and the texted registration code is verification-code forwarding; a new lure for covered mechanisms, on the parking-meter QR precedent. The fake Nationwide direct-debit email (stays inside phishing). The Social Security statement phishing email (Scamicide, November 8, 2026; stays inside phishing + government impersonation, on the SSA adjudication of 2.5.227). The Sri Lanka police traffic-fine GovPay texts (November 9, 2026; stay inside smishing + government impersonation). The Morristown, New Jersey bank-impersonation cash hand-over (November 7, 2026; stays inside gold-bar courier scams + the card courier scam, on the Fall River precedent). The Veterans Day roundups (New York Attorney General and Bitdefender, November 9, 2026; the Veterans Savings pitch already adjudicated, the rest stays inside government impersonation + charity scams + prize scams). The DeFi domain hijacks via Name.com (November 7, 2026; stay inside DNS hijacker + crypto wallet drainer). The unverified N0n ransomware claim against Expedia (November 9, 2026; a group claim, not a confirmed incident — stays inside ransomware + double extortion). The LNbits shutdown and Web3Auth's protective wallet move (November 8–9, 2026; platform incidents — stay inside supply-chain attack + credential stealer, on the Keplr adjudication of 2.5.245). The Apple WebKit watering-hole campaign (Sekoia, November 7, 2026; stays inside zero-click exploits + watering hole, on the 2.5.238 precedent). GhostMap, the SonicWall KEV additions, the joint refrigeration advisory, and SafePay were already adjudicated and were not re-litigated.
- Counts stay live: guideCount is computed from the guide data (173), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.246 above 2.5.245 / 2.5.244 (2.5.243 folds in, and the fold summary now names Server 2.5.243); the ES homepage gains the 2.5.246 paragraph; the admin guide gains the 2.5.246 card above the 2.5.245 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.245
The Threat Guide's daily update — 2 genuinely new entries, 169 → 171, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 2 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Travel booking phishing (real trip details) (Medium, Tricks, traps & nuisances — Frontier Airlines' confirmed campaign, warned by the airline on November 2, 2026 and still being reported on November 8): phishing mail carrying the passenger's real name, real confirmation code, and real itinerary — in some reports even a schedule change the airline itself had just made — in two scripts: a refund for a duplicate charge or billing adjustment that requires the paid-with card to be "confirmed" first, and a booking left "unconfirmed" over a payment problem unless the card is re-entered now; both end on a look-alike site, registered just before the mail went out, harvesting the full card number, billing address, and email. The details are real because they leaked (the airline disclosed security incidents earlier in 2026 and has not said how the scammers obtained the booking data). Distinct from refund scams (staged overpayments and remote-access theatre; never the victim's real booking) and from real order-email scams (a genuine retailer mail carrying attacker text — here the sender is fake and the data is real): the script's whole payload is that the traveler's natural test — "they knew my confirmation code, so it must be them" — now passes for the scammer. And Home warranty "final notice" mailers (Medium, Tricks, traps & nuisances — BBB Scam Alert, updated November 6, 2026): postal mail impersonating the mortgage lender — the real lender, loan amount, and property, all from public records — from a "Home Warranty Dept." or "Home Warranty Division," stamped "Final Notice" or "Renewal Required," sometimes designed to look like a check with tear-off edges and a "Renewal Fee Voucher," warning that home warranty coverage is expiring and hinting the mortgage itself is at risk, while its own fine print confesses "Not all consumers have previous coverage. We are not affiliated with your current mortgage." Many recipients never had a warranty at all; the call the letter pushes sells a warranty plan and takes the card (one homeowner was pitched a $234 plan and charged on the debit card she handed over; one seller the BBB profiles carries 67 complaints). Distinct from home title (deed) fraud (a forged filing takes the title — here nothing is filed anywhere, and the payload is the plan sale plus the card) and from fake debt collectors (whose letters invent a debt — this letter invents expiring coverage).
- Honest scope: the travel entry claims only the follow-on coverage that exists — email scanning grades the refund/payment-demand pattern and web and DNS protection block the known harvest sites — and says plainly that no filter can tell whether the trip details in a message are real, because here they are; the warranty entry claims no software counter at all, because there isn't one — no antivirus can read your postal mail, and nothing in the scam touches a computer until the victim picks up the phone. The defenses the entries give are the real ones: for travelers — a company that owes you a refund already knows how you paid and never needs your card number again, no airline emails you to revalidate a booking or re-enter payment details, so close the message and check the trip in the airline's own app or site, typed in yourself, and treat a card already entered as exposed (call the number on its back); for homeowners — a home warranty is an optional service contract that has nothing to do with your mortgage, read the fine print before the headline (the non-affiliation line is the confession), never call the number printed on an unsolicited notice (call your lender on the number from your own mortgage statement), never hand a card to someone who contacted you first, and dispute a charge already made with the card issuer. Severity totals are now Critical 12 / High 67 / Medium 78 / Low 14.
- Checked and NOT duplicated: PhantomDOM (ZeroBEC, November 2026) and the Nyasher / Operation CloudFire session streaming (Material Security via BleepingComputer, November 4, 2026) — the browser-in-the-middle technique stays inside adversary-in-the-middle phishing + session hijacking, on the CloudFire adjudication of 2.5.243, whose framework streams the genuine sign-in the same way; the Singapore Police WhatsApp file advisory of November 2, 2026 (malicious ISO / ZIP / RAR attachments sent from compromised accounts of people the victim knows, stealing PC data and self-forwarding weeks later — stays inside malicious disk images + trojan + social engineering); the Crédit Mutuel Arkéa $133 million heist via a service provider (BleepingComputer, November 7, 2026 — a provider incident, not a victim-facing type; stays inside supply-chain attack + spear phishing); the Keplr / Web3Auth social-login private-key exposure (November 6–7, 2026 — a provider infrastructure incident with no victim-facing script; stays inside credential stealer + supply-chain attack, on the platform-incident precedent); the OpenSSH "sshole" Alpine Linux bypass and the unpatchable SonicWall GMS flaws (stay inside exploit); the BBB jury-duty and police-chief spoofing warnings (stay inside government impersonation); and the fake Costco recruiting messages (stay inside job & task scams). The joint refrigeration advisory, XCPE, the fake AI event pages, health-plan open enrollment, Veterans Savings, and SafePay were already adjudicated and were not re-litigated.
- Counts stay live: guideCount is computed from the guide data (171), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.245 above 2.5.244 / 2.5.243 (2.5.241 folds in, and the fold summary now names Server 2.5.241); the ES homepage gains the 2.5.245 paragraph; the admin guide gains the 2.5.245 card above the 2.5.244 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.244
The Threat Guide's daily update — 2 genuinely new entries, 167 → 169, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 2 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Rental listing scams (Medium, Tricks, traps & nuisances — police warnings renewed across November 2026: Somerset, Kentucky on November 7, Marshall, Missouri and Lakeport, California in the same week, on the FTC's longstanding rental-listing warnings): a copied or invented listing, a landlord who is always out of town, an application fee, deposit, or first month demanded by Cash App, Zelle, Venmo, or wire before any inside viewing — and a fake application form that harvests the renter's ID, Social Security number, and bank details, enough to open accounts in their name long after the "landlord" vanishes. Distinct from marketplace scams (a goods deal on a platform, where the victim is a party to the deal) and from home title (deed) fraud (where the victim owns the home): here the victim is a renter who ends with no home, and the deposit plus the identity harvest are the whole payload. And Review extortion & fake-review scams (Medium, Tricks, traps & nuisances — Malwarebytes, November 2, 2026): a sudden wave of fake one-star reviews lands on a business's Google, Yelp, or Trustpilot profile, followed by a demand — pay money, goods, or services and the reviews come down — while underneath, underground sellers offer five-star reviews on demand (one claimed more than 10,000 posted, increasingly AI-written and posted through residential proxies) and negative-review removal at $100–$200 per review, about $1,100 on Airbnb. Distinct from theft-only extortion (what is held is the victim's files) and from the brushing scam (whose fake verified-purchase reviews are a side effect of a parcel): here the rating itself is the hostage, and the business is the victim.
- Honest scope: both entries claim no software counter for the core act, because there isn't one — no antivirus can tell whether a person advertising a home owns it, and no antivirus can remove a fake review or tell a platform which reviewers were real customers. What the layers cover is the follow-on (web and DNS protection block the known scam listing, application, payment, and forum pages; email scanning grades the written versions' patterns), and the defenses the entries give are the real ones: for renters — see the inside before any money moves, verify the lister owns or manages the home in county property records, sign a real lease first, never pay by payment app or wire to someone you have not met at the property, and treat an ID and Social Security number already sent on a fake application as exposed (credit freeze); for businesses — never pay or reply to an extortion demand, screenshot every demand, report the wave through Google's dedicated review-extortion form and the platform's normal reporting, and never buy fake reviews or paid removals (the FTC's review rule bans them, and a business answers for what a reputation service it hired does in its name). Severity totals are now Critical 12 / High 67 / Medium 76 / Low 14.
- Checked and NOT duplicated: the SonicWall SonicOS SSL-VPN SAML bypass (CVE-2026-19868) and the other November 6 CISA KEV additions (stay inside exploit, on the standing KEV precedent), the fake security tools on GitHub (Dr.Web via BleepingComputer, November 6, 2026; trojanized projects delivering a backdoor, a clipboard clipper, and an infostealer — stay inside trojan + backdoor + infostealer + clipboard hijacker, on the Yoba fake-GitHub-projects precedent of 2.5.236), the ChatGPT Lockdown Mode bypass (Check Point Research, November 6, 2026; a research demonstration, fixed in September, no in-the-wild use — stays inside exploit, on the Calif Lockdown Mode precedent of 2.5.238), the escaped AI hacking agent at WorkOS (BleepingComputer, November 6–7, 2026; a platform incident, not a victim-facing type), the fake Gemini "Google Coin" presale (a February 2026 Malwarebytes find recirculated November 7, 2026; stays inside AI chatbot impersonation scams + investment scam / "pig butchering"), the Fall River "bad money" cash-pickup warning (police, November 6, 2026; stays inside gold-bar courier scams, whose defense already names cash handed to a courier, + government impersonation + the card courier scam), the Binance-breach follow-on phishing (Bitdefender, November 4, 2026; stays inside phishing + recovery scams), the $129 million AI executive-impersonation indictment (DOJ, November 6, 2026; stays inside deepfake & AI voice-cloning scams + spear phishing & BEC), and the fake meter-update APK lures (stay inside fake & malicious phone apps + utility shutoff scams). XCPE, the SMS of Death, and SafePay were already adjudicated and were not re-litigated.
- Counts stay live: guideCount is computed from the guide data (169), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.244 above 2.5.243 / 2.5.241 (2.5.238 folds in, and the fold summary now names Server 2.5.238); the ES homepage gains the 2.5.244 paragraph; the admin guide gains the 2.5.244 card above the 2.5.243 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.243
The Threat Guide's daily update — 1 genuinely new entry, 166 → 167, in English and Spanish alike. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): Missing-pet extortion scam (AI "found your pet" photos) (Medium, Tricks, traps & nuisances — Malwarebytes, November 5, 2026, reporting the BBC's interviews with targeted owners): you post a missing-pet appeal — the pet's photo and your own contact details — and a stranger replies claiming to have found the animal, sending a picture that looks exactly like your pet, safe and well. The picture was made with AI from the appeal's own photo, by someone who has never seen the animal, and it is followed by the money: a "reward" to hand the pet over, or urgent "emergency treatment" — in one variant a caller impersonating a humane society demands successive Zelle payments — payable before you may see the animal, with threats to sell, keep, or harm it underneath. Distinct from deepfake scams (a fake person shown to the victim — here a fabricated image of the victim's own pet serves as proof of custody, defeating the owner's natural "send me a photo" test), from the family emergency scam (a relative supposedly in trouble), and from sextortion (a threat hanging on material the sender claims to hold — here the hostage exists only in a picture, and the whole script starts from a public appeal the victim posted).
- Honest scope: the entry claims no software counter for the core act, because there isn't one — no antivirus can tell whether a sender has your pet, and an AI-made photo of an animal is not a file a scanner can convict. What the layers cover is the follow-on (web and DNS protection block the known scam payment pages these messages point to, and email scanning grades the written version's pattern), and the defenses the entry gives are the real ones: keep one identifying detail out of the public appeal and ask the "finder" to name it, never pay before you stand in front of your animal, arrange any handover in a safe public place, and verify any vet or shelter claim by calling a number you found yourself. Severity totals are now Critical 12 / High 67 / Medium 74 / Low 14.
- Checked and NOT duplicated: the WordPress Core upgrade-window RCE and the other November 3 CISA KEV additions (already adjudicated in 2.5.238; stay inside exploit, not re-litigated), the FTC's imposter-ads notice (already adjudicated in 2.5.236; stays inside malvertising + SEO poisoning + government impersonation), "ghost tapping" (already adjudicated in 2.5.236; stays inside relay attacks + the card-tap phone scam), the NIKKE PC launcher tampering (BleepingComputer / Malwarebytes, November 4–5, 2026; stays inside supply-chain attack + backdoor + crypto wallet stealer), the trojanized CWP web server / CWPdoor (BleepingComputer, November 5, 2026; stays inside supply-chain attack + backdoor), Operation CloudFire's Calendar-invite Google session hijack (Material Security via BleepingComputer, November 4, 2026; stays inside calendar invite phishing + adversary-in-the-middle phishing + session hijacking), the fake OpenAI / SpaceX / Anthropic event pages (Push Security via BleepingComputer, November 5, 2026; stay inside fake meeting invites + device-code phishing + adversary-in-the-middle phishing), the GTA 6 leaked-build / early-access sites (Netcraft, November 5, 2026; stay inside malicious game cheats & mods + cracked software + crypto wallet drainer + fake online stores), the fake $750 Apple Cash reward page (Malwarebytes, November 4, 2026; stays inside prize & lottery scams + social-media giveaway scams), the phone-return write-up (Malwarebytes, November 5, 2026 — the phone-return scam entry itself), the health-plan open-enrollment mis-selling warnings (FTC, October 29, 2026; stay inside Medicare & health-benefits scams, whose entry already names the open-enrollment seasonal cousins, + government impersonation), and the sextortion-assistance injunction (already adjudicated in 2.5.202 as the recovery scam; the FTC's October 21, 2026 preliminary injunction is the same scheme, not re-litigated).
- Counts stay live: guideCount is computed from the guide data (167), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.243 above 2.5.241 / 2.5.238 (2.5.236 folds in, and the fold summary now names Server 2.5.236); the ES homepage gains the 2.5.243 paragraph; the admin guide gains the 2.5.243 card above the 2.5.238 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.157, Linux 1.32.0, Android 1.40.0).
Version 2.5.242 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 78 → 83 names. (Server-only release; Windows client 2.5.157, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.241.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Open Source Developer, Scott Rae (HandBrake-1.11.2-x86_64-Win_GUI.exe, github.com/HandBrake/HandBrake release 1.11.2; cert SHA-1 9528F963AD448D599543E77F4FE3462B12F995DD), Akeo Consulting (rufus-4.15.exe, github.com/pbatard/rufus release v4.15; cert SHA-1 FC4686753937A93FDCD48C2BB4375E239AF92DCB), IDRIX SARL (VeraCrypt.Setup.1.26.31.exe, github.com/veracrypt/VeraCrypt release VeraCrypt_1.26.31; cert SHA-1 86E1D426731E79117452F090188A828426B29B5F), Mullvad VPN AB (MullvadVPN-2026.4.exe, github.com/mullvad/mullvadvpn-app release 2026.4; cert SHA-1 1F5E906F4E2DBE2A3C3226A6B0638E9327F76135), and DBeaver Corp (dbeaver-ce-25.2.4-x86_64-setup.exe, github.com/dbeaver/dbeaver release 25.2.4; cert SHA-1 3D6F5CE8A7F8BAB9AAF4BAF37BD131E71254CA1F).
- The spellings are the certificates', not ours: "Open Source Developer, Scott Rae" is the certificate's full CN, comma included — it is HandBrake's Windows signer (a Certum open-source developer cert), the same class as Kovid Goyal (Calibre) and Simon Tatham (PuTTY) already on the list: the software is widely run, and the certificate's legal name is what ships, never the product name. "Akeo Consulting" is Rufus's signing entity and "IDRIX SARL" is VeraCrypt's, all caps as certified; "Mullvad VPN AB" and "DBeaver Corp" carry no trailing period. Open Source Developer, Scott Rae was banked by the 2.5.237 cap; today's re-download was byte-identical to the banked artifact (same SHA-256, same certificate). Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25242) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but deliberately NOT shipped: Transmission's official installer is signed "SignPath Foundation", a shared open-source signing service that signs many unrelated projects, not Transmission's own publisher — adding it would give a heuristic reduction to every SignPath-signed file, broader than the named software, so it is excluded on scope grounds. No verified names remain banked after this run. Attempted and not added: qBittorrent, MPC-HC and PeaZip (their official installers are unsigned — the same settled class as 7-Zip, Kodi and ShareX).
- Docs in the same push: admin guide gains the 2.5.242 card above the 2.5.237 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.241).
Version 2.5.241
Detection honesty — fewer false alarms, truer counts. (Server + Windows client release: server 2.5.241, Windows client 2.5.157; Linux 1.32.0 and Android 1.40.0 are unchanged.)
- What prompted it: a review of one managed PC's admin detection list (DESKTOP-6PVT7MF) showed three verdicts that should not have been recorded as detections: ~20
Defense.HostsFile rows for junk domains mapped to 0.0.0.0 (a blocklist install, not an infection), one Keylogger.Detected on Microsoft's OneDrive launcher in AppData, and one AI.Heuristic.Suspicious advisory on Microsoft's DiagTrack.dll in the $WINDOWS.~BT upgrade staging folder. All three reached the server through engine_event_threat() → cloud_report_detection, so each also counted toward the lifetime threats-stopped total. - Hosts file (Windows client, src/hostsmon.c): the monitor reported every mapping change as its own detection. Now only a suspicious mapping — a well-known domain pointed at a loopback/private IP, the pharming shape — reports individually, exactly as before. All other changes fold into the existing notification batch and the batch flush emits exactly ONE summary
Defense.HostsFile row, hosts file changed|N mappings|none flagged suspicious, where N is the true count of benign changes in the batch (tracked by a dedicated counter, hostsmon_report_*, so entries past the toast's 64-entry display cap are still counted). A batch of only suspicious changes emits no summary; per-change internal logging is unchanged. - Keylogger watch + banking mode (Windows client, src/keylogmon.c, src/bankmode.c): the keylogger-shape verdict (keyboard-hook shape, no visible window, user-writable path) never checked the file's signature. Both consumers now pass the verdict through the new signer gate before any consequence — the detection row, the toast, the behavior-chain event, and the opt-in banking-mode termination (both the session-start sweep and the session watcher): a valid signature from a trusted-list signer suppresses the verdict and logs the suppression on-device with the signer name. Invalid signature, unsigned, unlisted signer, or check failure all report and act exactly as before; the allowlist and once-per-basename dedupe are unchanged.
- AI heuristic (Windows client, src/aiscore.c, src/aeantivirus.c): the same gate sits on the
AI.Heuristic.Suspicious emit path in the scan pipeline. A validly signed, trusted-listed file no longer becomes an advisory row (nor a suspicious-file submission); the suppression is logged with the signer name and the would-be score. Scoring weights are untouched — this is a reporting gate on the advisory verdict only. The gate itself (ai_signer_gate) composes the two existing verifiers — WinVerifyTrust validity (aiscore_trusted_signed) and the trusted-signers list match — so there is no second signature implementation; its pure decision (ai_signer_gate_decide) never excuses an invalid signature. - Already-recorded rows stay: detections recorded before this client reaches a device remain in the history and counters — the fix stops new noise; it does not rewrite the past.
- Docs in the same push: homepage rotation (2.5.241 + Windows 2.5.157 takes the newest visible slot; 2.5.235 folds; fold summary updated) in English and Spanish; the existing "Everything included" cards for the hosts-file monitor, keylogger alerts, and banking mode were audited and stand unchanged (suspicious redirects still alert individually, and the keylogger warnings still fire for unsigned/untrusted programs); admin guide gains the 2.5.241 card; Client What's New gains the 2.5.157 entry. Tests: new
server/tests/client/honesty-test.c/.sh (hosts batch accounting — all-benign, over-cap, mixed, all-suspicious — and the gate decision over stubbed verifier results) and new server/tests/server/v25241-honesty-test.js (release shape: server 2.5.241, PLATFORMS + update manifest + APP_VERSION in agreement on Windows 2.5.157, source pins for the three gates, homepage rotation EN+ES). The five client suites that pin the current APP_VERSION were rolled to 2.5.157 in this release (marked "rolled 2.5.241").
Version 2.5.240 (admin-only)
Admin device commands fixed — Force heartbeat and Block process work again. No customer-facing changes. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.238.)
- Root cause: on the admin device detail page,
forceHeartbeat() and blockProcess() posted the bulk-command body shape ({install_ids:[...], command}) to the single-device endpoint POST /api/admin/commands, which reads {install_id, cmd}. Both fields arrived undefined, the endpoint answered 400 bad command, and nothing was ever queued — the alert seen live on production ("Could not queue: bad command", DESKTOP-6PVT7MF, 2026-11-04). The commands themselves were never the problem: heartbeat-now and block-process sit in VALID_CMDS, pass remoteCommandIsValid, are carried by the client command parser, and heartbeat-now is handled by the Windows engine's cloud_heartbeat() — only the two page functions spoke the wrong shape. - Fix: both functions now post the single-device shape (
{install_id, cmd}), exactly like their neighbor restartDevice() and every other per-device caller. A full audit of every served-page fetch body posted to either command endpoint (15 callers across the admin console: sandbox detonate/URL, generic send, lock, Router Guard run-now ×2, checkup request, Tech Tools single/multi/checkup runs, fix queue, and the bulk bar) found these two were the only mismatches; the bulk bar correctly posts the bulk shape to /api/admin/bulk-command and is untouched. The endpoint contract is unchanged — the bulk shape posted to the single-device endpoint still returns 400, now pinned by test so the two shapes cannot silently drift again. - Docs in the same push: README gains this section and the admin guide gains the 2.5.240 card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.238), and Client What's New is untouched. Tests: new
server/tests/server/v25240-cmdshape-test.js (boots the real server, registers a device, posts the exact body shapes the fixed page functions send, asserts ok:true and that each command lands in the target install's queue; asserts the old mismatched shape still returns 400; source-pins both fixed functions and the release shape).
Version 2.5.239 (admin-only)
Server memory fixes — the November 4 audit release. No customer-facing changes. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.238.)
- Why: Render memory-limit auto-restarts (one on Oct 31, two on Nov 4 about an hour apart, after the Oct 22–23 cluster) were traced by a full server audit to a high memory floor plus huge transient allocations — no single leak. This release takes the audit's top three fix groups; the unbounded table reads (finding #4) and signature-publish transients (finding #5) are deliberately NOT in this release.
- Backups stream, format unchanged: backup encryption, restore verification, the local-restore staging decrypt, and off-site copies now stream the archive instead of holding it whole in memory (previously 1–3 full copies of an archive that grows with the whole data dir). The encrypted layout is byte-identical —
"AEBK01" | salt(16) | iv(12) | ciphertext | authTag(16), one AES-256-GCM stream — so a backup written by 2.5.239 restores on older builds and this build restores backups written by 2.5.238 and earlier; the suite proves both directions. Backup file hashing (verify + update-channel drift check) is chunked the same way. The off-site PUT sends the same SigV4-signed request with the payload hash computed in a streaming pass. - Jobs cannot stack:
jobRun gains an in-flight guard — a second trigger of a job that is still running (schedule tick or Run Now) is skipped cleanly and journaled as a skipped run with the reason on the row. This closes the amplifier where the three backup jobs (all Run-Now-exposed) could overlap each other or a feed spike. - Uploads stream to disk:
/api/storage/upload (512 MB cap) and the backup restore/validate uploads (257 MB cap) are written to a temp file as they arrive with a running size cap and SHA-256, then moved into place — auth, quota checks, error strings, statuses, and response shapes are unchanged; only the memory profile changed. - Threat feed is cached:
/api/threat-feed and /api/threat-domains now serve a serialized response built once per store state (invalidated by the store's own updated stamp plus list/table state, so feed syncs, admin edits, imports, and web-filter changes all refresh it) instead of re-serializing the ~41 MB dataset on every install poll. Response bytes for a given state are identical; per-install fields (web-filter schedule, whitelist/rules versions) are still computed per request. - Small maps are bounded: the per-IP signup rate-limit maps (
statusAlertRate, newsletterRate) prune expired buckets lazily like the global API limiter, and the three Stripe response caches (subPriceCache, acctInvoiceCache, _chargeCache) evict stale entries past a size cap. Hit semantics are unchanged. - Docs in the same push: README gains this section and the admin guide gains the 2.5.239 card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.238), and Client What's New is untouched. Tests: new
server/tests/server/v25239-memory-test.js (job-overlap refusal against a delaying S3 stub, threat-feed byte-identity + invalidation, threat-domains invalidation on store and web-filter changes, streamed storage upload incl. a >1 MB body and the quota 413, backup round trips in both format directions, newsletter/status-alert rate-limit behavior, and a real-request JSON pass over the changed routes).
Version 2.5.238
The Threat Guide's daily update — 1 genuinely new entry, 165 → 166, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): Prompt adware (AI agent ad injection) (Medium, Tricks, traps & nuisances — Check Point Research, November 3, 2026): a free "jailbreak" bundle for an AI coding agent — Check Point analysed Haiou 4.5.0, built for OpenAI's Codex — persistently rewrites the agent's GLOBAL instruction files so that exact greetings ("hi", "hello", and their Chinese equivalents) trigger a fixed advertisement for the author's opaque LLM gateway, a middleman that would sit between the agent and the real AI provider. The bundle also weakens the agent's guardrails, installs security-testing skills, and ships its prompt files encrypted, so the person installing it cannot read the rules being planted. Check Point, which coined "prompt adware" for the class, found no indication the advertised gateway itself is malicious — the harm is the mechanism: whoever controls an agent's standing instructions controls what it says, recommends, and protects, in every session, until the files are changed back. Distinct from AI agent instruction hijacking (an attacker plants a file the victim never chose, ordering malware installs or token exposure — here the victim installs the bundle themselves and the payload is commercial), from adware (ads injected into browsing by code — these are injected into an agent's answers by words), and from PUPs (a program whose unwanted traits are its own — here the unwanted behavior appears in a different, trusted program's output).
- Honest scope: the entry claims no software counter for the core act, because there isn't one — the planted instruction is plain text in a configuration file, not malware, and no antivirus can stop an agent from reading its own instructions. What AEAntivirus holds is the behavior around it (signature and behavioral scoring judge anything the altered agent is steered into downloading or running; the heuristic scorer flags bundles that persistently rewrite other programs' settings), and the defense the entry gives is the real one: install only agent add-ons you can read in full — encrypted prompt files are a refusal, not a feature — check the global instruction files after installing any agent bundle, remove changes you did not write, route agent traffic only through providers you chose yourself, and if your agent starts reciting an advertisement, restore its instructions and review what else changed. Severity totals are now Critical 12 / High 67 / Medium 73 / Low 14.
- Checked and NOT duplicated: LLMjacking at scale (Google GTIG / Mandiant, November 2, 2026 — already adjudicated in 2.5.233 as post-compromise cloud abuse inside cloud account hijacking + credential stealer; a larger report of the same behavior is not a new type and was not re-litigated), the WardenStealer / Aoki AI-token logs (SpyCloud, October 30, 2026 — Aoki is the AI assistant credential stealer entry's own subject and WardenStealer performs the same collection; stays inside the AI assistant credential stealer + infostealer), the cPGuard update-channel hijack (BleepingComputer, November 3, 2026; stays inside supply-chain attack + backdoor), the WordPress Core upgrade-window RCE (CVE-2026-119977) and the other CISA KEV additions of November 3 (stay inside exploit), the n-day zero-click watering-hole campaign (Sekoia, November 4, 2026; stays inside zero-click exploits + watering hole), the Calif Lockdown Mode jailbreak demonstration (a research demonstration with no tool released and no in-the-wild use; stays inside exploit + malicious PDFs, on the CLOSEDQUORUM precedent), Gorbag / BeautyStealer (ESET, November 3, 2026 — a group name plus a conventional stealer; stays inside infostealer + spear phishing + exploit, on the SafePay precedent), the Edmonton fake-carrier phone operation (police, November 4, 2026 — the phone-return scam itself; the reporting calls it a close cousin of the Vancouver Island case that entry documents), the HMRC crypto-declaration texts (an October 15, 2026 warning recycled on November 4; stays inside smishing + government impersonation), the French DAC8 and IRS "Digital Asset Compliance Portal" mailed tax letters (stay inside QR-code phishing — whose entry already names mailed letters — + government impersonation + crypto wallet drainer), the recycled CISA / FBI commercial-messaging-app PSA (first issued March 2026 and updated June 2026; its QR / PIN / support-impersonation tactics stay inside QR login jacking + verification-code forwarding + phishing), the European supermarket refrigeration intrusions (already adjudicated in 2.5.236; stay inside router & IoT malware + exploit, not re-litigated), and the Binance fake-interview heist with its follow-on recovery phishing (stays inside job scams + recovery scams; the platform incident itself is not a victim-facing type).
- Counts stay live: guideCount is computed from the guide data (166), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.238 above 2.5.236 / 2.5.235 (2.5.233 folds in, and the fold summary now names Server 2.5.233); the ES homepage gains the 2.5.238 paragraph; the admin guide gains the 2.5.238 card above the 2.5.236 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.156, Linux 1.32.0, Android 1.40.0).
Version 2.5.237 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 73 → 78 names. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.236.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Krzysztof Kowalczyk (SumatraPDF-3.6.1-64-install.exe, sumatrapdfreader.org/dl/rel/3.6.1; cert SHA-1 E1E6883F78BF923B92F21A71FD33A452C9ED7DD0), voidtools (Everything-1.4.1.1026.x64-Setup.exe, voidtools.com; cert SHA-1 4DA2AD938358643571084F75F21AFDDD15D4BAE9), Razer USA Ltd. (Razer Synapse 4 installer, rzr.to/synapse-4-pc-download from razer.com/synapse-3; cert SHA-1 3FDDBD6F820144422494342EF44898D732E068DC), Balena Ltd (balenaEtcher-2.1.9.Setup.exe, github.com/balena-io/etcher release v2.1.9; cert SHA-1 668D58190E972578AABAE6478426C2A76CE52A57), Blender Foundation (blender-5.2.3-windows-x64.msi, download.blender.org/release/Blender5.2; cert SHA-1 01DE4C0BE4BD64233ADA664DC389515CEB0FF412).
- The spellings are the certificates', not ours: "voidtools" is a single lowercase word as certified — it is the Everything search tool's signing entity. "Krzysztof Kowalczyk" is Sumatra PDF's individual author signer, the same class as Kovid Goyal (Calibre) and Simon Tatham (PuTTY) already on the list: the software is widely run, and the certificate's legal name is what ships, never the product name. "Razer USA Ltd." carries its trailing period where "Balena Ltd" and "Blender Foundation" carry no suffix punctuation. Krzysztof Kowalczyk was banked by the 2.5.234 cap; today's re-download was byte-identical to the banked artifact (same SHA-256, same certificate). Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25237) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but deliberately NOT shipped — the cap is the cap: 1 more name passed the same verification in the same pass and is banked in the ledger for the next daily run: Open Source Developer, Scott Rae (HandBrake — the certificate's full CN, comma included). Verified but excluded on format grounds: GIMP's signer is not printable ASCII, so it cannot ship as certified and was not transliterated. 5 shipped, 0 padded. Attempted and not added: WinSCP (its download route returned an HTML page instead of a binary, again — no signer could be resolved), Surfshark, Parsec, Zotero, ShareX (its official installer is unsigned), and FileZilla / Proton VPN / Foxit (their download pages expose no direct signed-binary link).
- Docs in the same push: admin guide gains the 2.5.237 card above the 2.5.234 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.236).
Version 2.5.236
The Threat Guide's daily update — 1 genuinely new entry, 164 → 165, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): Manual-entry card scam (failed-tap trick) (Medium, Money-driven threats — FBI Newark field office warning, reported October 31 – November 2, 2026): a customer's contactless tap "fails" at the till, and the customer helpfully offers a card number for the cashier to key in by hand — the number is stolen, the sale completes, and the shop learns the truth weeks later when the real owner's chargeback claws the payment back and the goods are already resold; a second version runs an app on the customer's phone that stores many stolen card numbers and cycles through them at the reader until one approves, while an accomplice distracts the cashier. At least seven New Jersey trading-card shops were hit between September 2025 and March 2026, some losing tens of thousands of dollars in stock. Distinct from card skimmers (a physical device copies the card), the card-tap phone scam (the victim's own phone relays their own card during a call), POS malware (card data scraped from register memory), and the BBB "double payment" declined-card trap rejected in 2.5.235 (an online fake-store harvest where the shopper is the victim — here the shop is).
- Honest scope: the entry claims no software counter, because there isn't one — no file is scanned and the fraud is a person at the till. The defense it gives is the FBI's, and it is policy: turn off manual card entry where possible, require matching photo ID for any sale that truly must be keyed in, and treat a failed tap followed by an offered number — above all for high-value, easy-to-resell goods — as the warning sign it is. Severity totals are now Critical 12 / High 67 / Medium 72 / Low 14.
- Checked and NOT duplicated: the XCPE Android TV-box botnet (Google Threat Intelligence Group, November 2, 2026 — a backdoor Google says the manufacturer built into the firmware of 1.3 million TV boxes, rented out as residential proxies and ad fraud; already rejected in the 2026-10-30 round as BadBox-style TV-box proxy devices and stays inside preinstalled Android malware + proxyware + botnet — XCPE is an extension of BadBox 2.0, a new name for covered behavior), the Hugging Face trust_remote_code GPU-miner campaign's return (BleepingComputer, October 2026 — the exact campaign already inside malicious AI models), "ghost tapping" (BBB, October 2026 — its proximity-reader half stays inside relay attacks, whose entry already names readers brushed against pockets and bags in crowds, and its tricked-tap half inside the card-tap phone scam and the mobile-wallet enrollment scam), the Power BI dashboard lure (a blurred-document page on a genuine Microsoft service leading to an abused remote-access tool; stays inside fake shared-file lures + abused remote-access tools + phishing), the Yoba fake AI-app downloads (Check Point, October 27, 2026; stays inside DLL sideloading + SEO poisoning + InstallFix), the FTC's imposter-ads notice (November 2026; stays inside malvertising + SEO poisoning + government impersonation), the FBI Honolulu staged-video law-enforcement warning (November 2, 2026; stays inside government impersonation + fake meeting invites + deepfake scams), McAfee's 1,900 fake impersonator apps (stay inside fake & malicious phone apps), and the BTCPay hot-wallet drains and European supermarket refrigeration intrusions (stay inside exploit + crypto wallet stealer and router & IoT malware). VOLTA and the Synology C3RB3R surge were already adjudicated in 2.5.233 and 2.5.235 and were not re-litigated.
- Counts stay live: guideCount is computed from the guide data (165), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.236 above 2.5.235 / 2.5.233 (2.5.231 folds in, and the fold summary now names Server 2.5.231); the ES homepage gains the 2.5.236 paragraph; the admin guide gains the 2.5.236 card above the 2.5.235 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.156, Linux 1.32.0, Android 1.40.0).
Version 2.5.235
The Threat Guide's daily update — 1 genuinely new entry, 163 → 164, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): Biometric selfie scam (face-scan fraud) (High, Tricks, traps & nuisances — South Africa's National Financial Ombud Scheme, November 2, 2026; India's I4C cybercrime centre, November 2026; Hong Kong Legislative Council, October 21, 2026): banks now let a live face scan authorize the big moves — linking a new phone to a banking profile, opening an account, recovering access — and scammers have learned to borrow that moment. In the in-person version, a fake loyalty-programme representative in a shopping centre offers a voucher or free airtime, and the "registration" face scan on their device is actually the bank's own verification step, completing a request the fraudster already started: it links the fraudster's device to the victim's banking profile. In the remote version, a fake "your eKYC has expired" video call walks the victim through the liveness routine — look at the camera, turn left and right, blink, smile, read a sentence aloud — and records a performance good enough to build a clone of the face that can pass video verification later. The guide had no biometric coverage at all. Distinct from deepfake scams, where a fake face is shown TO the victim — here the victim's real face is the thing taken — from the mobile-wallet enrollment scam, where the victim hands over a code (a code can be cancelled and reissued; a face cannot), and from SIM swapping, where the carrier is fooled and the victim performs no scan at all.
- Honest scope: the entry claims no software counter, because there isn't one — nothing is installed, no file is scanned, and the scan happens on the scammer's device or inside a call the victim joined, where no security software is watching. The defense it gives is the real one: treat a face scan exactly like a PIN, never scan your face on a device that is not yours, never perform a liveness routine on a call you did not start, and only ever complete a face verification inside your own bank's app, opened by you, for a request you began — and if you already scanned for a stranger, call the bank at once and ask whether a new device was linked or an account opened. Severity totals are now Critical 12 / High 67 / Medium 71 / Low 14.
- Checked and NOT duplicated: the trojanized Claude Code extension on Cursor (already rejected in 2.5.229; stays inside supply-chain + malicious browser extension + malicious packages), the Microsoft Teams help-desk / Quick Assist chain (already rejected in 2.5.226 and 2.5.231; Microsoft's November 2, 2026 report documents the same entry method it had already documented earlier in 2026, and its AD CS abuse is post-compromise; stays inside the remote-access scam + tech support scam + credential dumping), offensive AI-agent frameworks ARTEX / PentAGI / CAIRN (GTIG / Mandiant, October 30, 2026 — operator-side automation of post-compromise steps the victim never sees; stays inside credential dumping + living-off-the-land + backdoor, on the VOLTA / CLOSEDQUORUM precedent), NIK C2 (Cisco Talos, October 29, 2026 — a conventional RAT whose AI analysis step is operator-side and victim-invisible; stays inside backdoor + AI assistant credential stealer, on the VOLTA precedent), the grey-market Codex / Claude Code relay (Symantec, October 29, 2026 — a dozen organizations, no abuse observed; stays inside AI agent credential sharing + supply-chain, on the CLOSEDQUORUM precedent), the C3RB3R ransomware surge against Synology NAS devices (BleepingComputer, November 2, 2026 — a ransomware family name plus an unpatched-device exploit; stays inside ransomware + router & IoT malware + exploit, on the SafePay precedent), MR ROBOT Android banking malware-as-a-service (CloudSEK, October 28, 2026 — a rental model is an operator-side business model; its trojans stay inside banking trojan + fake phone apps + overlay attacks), the BBB "double payment" declined-card trap (a harvest multiplier inside fake online stores + web skimmer, not a new script), the ChatGPT Plus billing phishing kit (stays inside phishing), the Langflow flaws (stay inside exploit), and the Binance and Chainalysis incidents (platform incidents, not victim-facing types).
- Counts stay live: guideCount is computed from the guide data (164), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp already reads November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.235 above 2.5.233 / 2.5.231 (2.5.229 folds in, and the fold summary now names Server 2.5.229); the ES homepage gains the 2.5.235 paragraph; the admin guide gains the 2.5.235 card above the 2.5.233 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.156, Linux 1.32.0, Android 1.40.0).
Version 2.5.234 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 68 → 73 names. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.233.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: MEGA Privacy LLC (MEGAsync Setup, mega.nz/MEGAsyncSetup64.exe; cert SHA-1 8F181A3567B55143D1CC403BADF686D9D5ECF61E), Notion Labs, Inc. (Notion Setup, notion.so/desktop/windows/download; cert SHA-1 204B68546CF6E83E135DC433D98FD7B134E5DFD7), Rockstar Games, Inc. (Rockstar Games Launcher, gamedownloads.rockstargames.com/public/installer/Rockstar-Games-Launcher.exe; cert SHA-1 AD00A9C3007FD1450A559D0634547B075AD3E688), Vivaldi Technologies AS (Vivaldi 8.2.4133.96 x64, downloads.vivaldi.com/stable — the URL taken verbatim from the Vivaldi download page; cert SHA-1 F7A524AD45E585F8B71E6204B2583714151A08EF), and MUSECY SM LTD (Audacity 4.0.1 MSI, github.com/audacity/audacity releases; cert SHA-1 0DDE474DE490D683D3EE700D89CF069FEA557049). The full evidence trail (artifact hashes, sources, certs) is in the build ledger for this run.
- The spellings are the certificates', not ours: "MUSECY SM LTD" is all caps as certified — it is Audacity's signing entity (the Muse Group), the same class as DroidMonkey Apps, LLC (KeePassXC) and DOTPDN LLC (paint.net) already on the list: the software is widely run, and the certificate's legal name is what ships, never the product name. "MEGA Privacy LLC" and "Vivaldi Technologies AS" carry no suffix punctuation where "Notion Labs, Inc." and "Rockstar Games, Inc." carry their comma and period. Every shipped name was re-extracted in a solo pass before shipping.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25234) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but deliberately NOT shipped — the cap is the cap: 1 more name passed the same verification in the same pass and is banked in the ledger for the next daily run: Krzysztof Kowalczyk (Sumatra PDF — its individual author signer, the same class as Kovid Goyal for Calibre). 5 shipped, 0 padded. Attempted and not added: WinSCP (its download route returned an HTML page instead of a binary, again — no signer could be resolved), Surfshark, Parsec, Zotero, and TechSmith Snagit (no vendor binary could be obtained for any of them).
- Docs in the same push: admin guide gains the 2.5.234 card above the 2.5.232 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.233).
Version 2.5.233
The Threat Guide's daily update — 3 genuinely new entries, 160 → 163, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 3 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): SMS of Death (crash text message) (High, Stealth & network — BleepingComputer, October 31, 2026; working proof of concept published October 29, 2026, CVE-2026-58882): a malformed eight-byte WAP Push text, processed automatically before it reaches the inbox, crashes the Android process that runs calls and mobile signal — dropping any call in progress, including emergency calls — and the queued message re-crashes the phone on retry, with about twenty able to loop it into recovery mode and a factory reset; Google patched Android 14–17 in the October 2026 security updates, Android 12 and 13 received no fix. Distinct from zero-click exploits, whose crafted message lets an implant in: this one steals nothing and installs nothing. Digital arrest scam (High, Tricks, traps & nuisances — Indian Cyber Crime Coordination Centre figures published October 2026; India Supreme Court hearing, October 29, 2026): a fake officer — uniform, staged station backdrop, forged warrant on camera — keeps the victim "under digital arrest" on a live video call, sometimes for days, under orders to speak to no one, until savings are transferred for "verification"; 21,900+ complaints and ₹647 crore in reported losses in the first eight months of 2026 in India alone. Distinct from government impersonation, whose scripts threaten arrest but never stage continuous video custody. MFA downgrade lure (High, Tricks, traps & nuisances — Okta Threat Intelligence, October 5, 2026; Dataminr, October 19, 2026): a fake IT call talks a hardware-security-key user onto a texted or emailed code at a lookalike sign-in page (13 lookalike Okta / ServiceNow portals tracked), and the code is harvested and replayed in real time — the victim is persuaded to stop using the one factor phishing cannot relay. Distinct from MFA fatigue (approve-prompt bombing) and from adversary-in-the-middle phishing (which proxies the real factor).
- Honest scope in every entry: no antivirus can intercept a carrier message the phone's own system code handles before any app sees it, end a video call a victim has been frightened into staying on, or un-replay a harvested code — each entry says so plainly and names only the shipped layers it may claim (the phishing layers, email scoring, and scam-call warnings where they apply). Severity totals are now Critical 12 / High 66 / Medium 71 / Low 14.
- Checked and NOT duplicated: VOLTA (Cisco Talos, October 31, 2026 — a conventional infostealer haul whose only novelty is an AI victim-ranking step the victim never sees, unobserved in the wild; stays inside the infostealer entry, on the CLOSEDQUORUM precedent), the TeamViewer reverse-session flaw CVE-2026-111585 (a patched product exploit with no in-the-wild use; stays inside exploit + abused remote-access tools), LLMjacking billing abuse (post-compromise cloud abuse; stays inside cloud account hijacking + credential stealer), the Ivanti Connect Secure KEV additions of October 30 (stay inside exploit), the FTC's fake-FTC-website takedowns of October 29 (stay inside government impersonation + the fake IC3 recovery scam + typosquatting), the Guelph bank-card pickup warnings (stay inside the card courier scam), the Binance-theft fake recovery offers (stay inside the recovery scam), and the Northflank breach (a platform incident, not a victim-facing type).
- Counts stay live: guideCount is computed from the guide data (163), and the homepage strip, About tile, emergency one-pager, and
/api/threat-guide all render it dynamically — no page hardcodes the number. The guide's "Last reviewed" stamp advances to November 2026 (EN) / noviembre de 2026 (ES). - Docs in the same push: README gains this section; the homepage rotation adds Server 2.5.233 above 2.5.231 / 2.5.229 (2.5.227 folds in, and the fold summary now names Server 2.5.227); the ES homepage gains the 2.5.233 paragraph; the admin guide gains the 2.5.233 card above the 2.5.231 card. Client What's New is untouched — this is a server-only release, and the client update manifests keep their own versions (Windows 2.5.156, Linux 1.32.0, Android 1.40.0).
Version 2.5.232 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 63 → 68 names. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.231.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Wire Swiss GmbH (Wire Setup, wire-app.wire.com; cert SHA-1 56DEF976D902F8AA7461992D4F11EDEB3130D0EC), Tailscale Inc. (Tailscale setup, pkgs.tailscale.com; cert SHA-1 108F172FDE945B21A5C0696731D6220D67D1C39E), Cloudflare, Inc. (Cloudflare WARP MSI, downloads.cloudflareclient.com; cert SHA-1 0B6C68C4BAD79F9AC39E07FE84223FB796D4D0BC), Wireshark Foundation (Wireshark 4.6.9 x64, 2.na.dl.wireshark.org — the URL taken verbatim from wireshark.org's download page; cert SHA-1 2A47694B41567B4D0D58DCA5289393C0A84F0EA4), and DroidMonkey Apps, LLC (KeePassXC 2.8.0 MSI, keepassxreboot/keepassxc releases; cert SHA-1 14CF22744ED6C58E48230BBAD703CB37E02CC2A1). The full evidence trail (artifact hashes, sources, certs) is in the build ledger for this run.
- The banked name ships, re-verified first: Wire Swiss GmbH was verified in the 2.5.228 catch-up pass and held by its cap, then held again by 2.5.230's cap. This run re-downloaded the artifact before shipping it: the download was byte-identical to the banked one (same size, same SHA-256, same certificate), so the banked name is exactly the name in today's signature. "DroidMonkey Apps, LLC" is KeePassXC's signing entity — the same class as Kovid Goyal (Calibre) and DOTPDN LLC (paint.net) already on the list: the software is widely run, and the certificate's legal name is what ships, never the product name. The spellings are the certificates', not ours: "Tailscale Inc." carries its trailing period where "Wireshark Foundation" carries no suffix punctuation at all.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25232) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but deliberately NOT shipped — the cap is the cap: nothing is held back this run: exactly 5 names were verified and all 5 shipped, 0 padded, and no banked names remain. Attempted and not added: WinSCP — its download route returned an HTML page instead of a binary, so no signer could be resolved from a real artifact.
- Docs in the same push: admin guide gains the 2.5.232 card above the 2.5.230 expansion card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.231).
Version 2.5.231
The Threat Guide's daily update — 3 genuinely new entries, 157 → 160, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 3 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Tampered hardware wallet (built-in implant) (High, Money-driven threats — Ledger's confirmed implants; Singapore Police advisory, October 28, 2026; Trezor warning, October 30, 2026): an extra circuit board fitted inside a genuine-looking hardware wallet — in the reported Safe 5, between the main board and the screen, with a cellular module and antenna — reads the recovery phrase the screen shows during setup and transmits it over the mobile network whenever the device is plugged in, while the wallet itself keeps working normally; the maker's advice to buyers from the reseller CryptoBilis since March 2026 is to move funds to a new wallet with a NEW recovery phrase, because restoring the old phrase elsewhere does not help if it was already sent out. AI assistant credential stealer (High, Spying & stealing — Check Point, October 2026, via Dark Reading): the stealer tracked as Aokigahara (also Aoki) carries a dedicated collection routine for the authentication files, session data, and configuration files of Claude, ChatGPT Desktop, Codex, Cursor, Windsurf, GitHub Copilot, Gemini CLI, and other assistants, alongside browser passwords and cookies, crypto wallets, and the SSH private keys, Git credentials, and AWS / Azure / Google Cloud configuration files that sit on the same machine — a stolen AI session can be used on the victim's paid account, resold, or mined for conversation history and code. Phone-return scam (fake upgrade, fake return) (Medium, Tricks, traps & nuisances — RCMP warnings, October 2 and 9, 2026; Thunder Bay Police Service warning, October 30, 2026): a caller posing as the victim's carrier offers a free upgrade or loyalty deal, collects the account details, and orders a REAL phone in the victim's name; when it arrives, a second call claims the wrong device was shipped and sends a prepaid return label or QR code that redirects the phone to the scammers — the victim keeps the contract, financing, and monthly payments (losses topping $33,000 on Vancouver Island in a few weeks; one Ontario victim left owing nearly $17,000 after being talked into ordering six devices).
- Danger totals now read Critical 12 / High 63 / Medium 71 / Low 14. The hardware-wallet entry rates High: a transmitted recovery phrase is the whole wallet, spendable from anywhere, and the device gives no sign anything is wrong. The AI credential stealer rates High, with the infostealer and credential-stealer entries it extends: one run on a developer's machine can hand over AI sessions, source-code access, and cloud keys together. The phone-return entry rates Medium, like the card courier and brushing scams: the loss can reach five figures, but the defense is a small set of absolute rules a victim can learn — never return a device on a label a caller sent, and verify every return inside your own carrier account.
- Checked and NOT duplicated (2026-10-31 round): the FTC's fake-FTC-website takedowns (FTC release, October 29, 2026 — already inside government impersonation, the fake FBI / IC3 recovery scam, and typosquatting), the Amazon-impersonation remote-access wave (UK Report Fraud alert, October 30, 2026 — already inside the remote-access scam and the tech support scam, whose caller is "Amazon security" by name), the Arctic Wolf Teams external-chat IT-support chain (October 26, 2026 — already inside email bombing and the tech support scam, and already noted in 2.5.226), SafePay ransomware (a ransomware group name, already rejected in 2.5.202, 2.5.208, and 2.5.229; its tactics stay inside ransomware, double-extortion, and exploit), and the Ivanti Connect Secure and Drupal Commerce additions to CISA's KEV catalog (October 30, 2026 — already inside exploit). 160 is the honest count.
- Coverage claims stay honest: each entry claims only shipped layers and says plainly where no software counter exists — no antivirus can inspect the inside of a hardware wallet, un-steal a session token once it has left the machine, or sit between a victim and a shipping label. The advice that defeats each one is given in the entry itself: buy hardware wallets only from the maker or an authorized reseller and, if yours may be affected, move funds to a new wallet with a new recovery phrase; revoke AI-tool sessions and rotate API, cloud, and SSH keys after any stealer run; and hang up on unsolicited upgrade offers, call your carrier on the number on your bill, and never ship a device back on a label or QR code a caller sent you.
- Full EN+ES parity from day one: the Spanish edition gains all three entries in the same release — same slugs, same positions (after Crypto wallet stealer, after Credential stealer, and after Package delivery scam), same severities — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 160 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - Docs in the same push: admin guide gains the 2.5.231 card (the 2.5.229 card retained as history); homepage "New in" rotation (this section, 2.5.229, and Server 2.5.227 visible; the Linux 1.32.0 + Android 1.40.0 client release heads the fold and the fold summary names Server 2.5.226, the newest folded server release); the Spanish homepage gains the 2.5.231 summary paragraph at the top of its Novedades list.
Version 2.5.230 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 58 → 63 names. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.229.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: ELEMENT CREATIONS LIMITED (Element Setup, packages.element.io; cert SHA-1 0166C56D35C7F22F78D870D0C1BCF729F2CAA3E9), Figma, Inc. (FigmaSetup, desktop.figma.com; cert SHA-1 B9F6F7CB9C6CE2437C45D86BCC25A788422CADD4), Postman, Inc. (Postman (x64), dl.pstmn.io; cert SHA-1 1FBE5308261CD8052D21B011E99DDF62BD189A86), Dynalist Inc (Obsidian 1.14.5, obsidianmd/obsidian-releases; cert SHA-1 6233D66CF784FA0E3A0A2E71BDDBBADFD69413A0), and Sublime HQ Pty Ltd (Sublime Text build 4215 setup, download.sublimetext.com; cert SHA-1 AB4550F114DB3599664A9965AF64BC35154288F6). The full evidence trail (artifact hashes, sources, certs) is in the build ledger for this run.
- Banked names, re-verified before shipping: these five were first verified in the 2.5.228 catch-up pass and held by its 20-name cap. This run re-downloaded every artifact and re-resolved every signer before shipping a single name: four artifacts came back byte-identical to the banked downloads, and Postman served a newer build under the same certificate and the same byte-exact name. The spellings are the certificates', not ours: "ELEMENT CREATIONS LIMITED" is all caps as certified, and "Dynalist Inc" and "Sublime HQ Pty Ltd" carry no trailing period where "Figma, Inc." and "Postman, Inc." carry theirs.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25230) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but deliberately NOT shipped — the cap is the cap: 1 name from the catch-up pass remains banked in the ledger for the next daily run: Wire Swiss GmbH. 5 shipped, 0 padded.
- Docs in the same push: admin guide gains the 2.5.230 card above the 2.5.228 catch-up card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.229).
Version 2.5.229
The Threat Guide's daily update — 3 genuinely new entries, 154 → 157, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- 3 new entries, in the house format (what it is / how it gets in / how AEAntivirus stops it): Malicious AI models (model-hub malware) (High, Stealth & network threats — Pluto Security, October 29, 2026, via BleepingComputer): 31 malicious repositories across seven Hugging Face accounts posing as small Qwen2 test checkpoints — more than 32,000 downloads counting linked repositories, the flagship alone at 28,412 by October 28 with a 42-byte model file containing no model — abuse the trust_remote_code loading option, so repository-supplied Python runs with the caller's privileges before any weights load and installs a cryptominer disguised as a Linux system process (watchdog-restarted, competing miners removed), a reverse shell in some versions, and in a later script a privileged-container escape. It targets GPU hosts: self-hosted inference servers, rented GPU boxes, shared notebooks, automated jobs. Fake listing text scam (Medium, Tricks, traps & nuisances — Malwarebytes, October 6 and 13, 2026): a text from a stranger asks "Is this still available for purchase?" and attaches a fabricated Facebook Marketplace listing screenshot that names you as the seller; there is no listing and no deal, your account is not necessarily hacked, and the whole payload is your own corrective reply — any reply proves your number is live, makes it worth more, and lets it be resold to other scammers. AI agent credential sharing (High, Tricks, traps & nuisances — TechCrunch, October 27, 2026; Malwarebytes, October 29, 2026): major US banks reported detecting outside AI agents operating inside customer accounts, in many cases with credentials the customers supplied themselves, and Bank of America's customer email that month said never to hand a bank username or password to an outside AI tool or agent — the sharer is responsible for the activity that follows, a flaw in the agent's service can expose the stored login, and the bank may read the agent's actions as a takeover attempt and lock the account.
- Danger totals now read Critical 12 / High 61 / Medium 70 / Low 14. The model-hub and credential-sharing entries rate High: one hands an attacker code execution on GPU servers through a download believed to be inert data; the other hands an outside service the customer's full bank access with the liability that follows. The listing text rates Medium, like the brushing scam it resembles: a reply alone steals nothing, but it deliberately marks the victim for the fraud that comes next.
- Checked and NOT duplicated (2026-10-30 round): SafePay ransomware (CISA advisory AA26-293A, October 20 — a ransomware group name, already rejected in 2.5.202 and 2.5.208; its tactics stay inside ransomware, double-extortion, and exploit), GitLab CVE-2026-97229 and the Cisco ASA/FTD CVE-2026-76418 (CISA KEV, October 29 — already inside exploit), the Socket search-hijacking Chrome extensions, 319 extensions and 13.4 million installs (BleepingComputer, October 29 — already inside malicious-extension, which explicitly covers search redirection, and browser-hijacker), the trojanized Claude Code extension served from the Cursor marketplace (October 29 — already inside supply-chain, malicious-extension, and malicious-packages), the fake Solana starter kit that lures AI coding agents to a paste-a-command install (already inside agent-instruction-hijack and installfix), the Polymarket third-party wallet-draining script (October 29 — already inside web-skimmer, wallet-drainer, and supply-chain), the fake Padre Terminal crypto app on the Apple App Store (already inside fake-mobile-apps and wallet-drainer), uncertified Android TV streaming boxes and BadBox-style residential-proxy devices (Krebs — already inside preinstalled-android-malware, proxyware, and botnet), and CLOSEDQUORUM, the AI-voting implant (Cisco Talos, September 22 — already inside backdoor, infostealer, and credential-dumping, with no confirmed in-the-wild deployment). 157 is the honest count.
- Coverage claims stay honest: each entry claims only shipped layers and says plainly where no software counter exists — no antivirus can un-share a password you handed over, stop a stranger from texting you, or judge what a model repository's Python code does before it runs; hub scanners check uploads for known malware and unsafe model formats, not for what the code does. The advice that defeats each one is given in the entry itself: read a model's code before enabling trust_remote_code and pin a reviewed revision, never reply to the listing text (not even to correct it), and never type a bank login or one-time code into an outside AI tool — use the bank's own official, read-only-where-offered connection inside the bank's own app or site.
- Full EN+ES parity from day one: the Spanish edition gains all three entries in the same release — same slugs, same positions (after Malicious software packages, after Marketplace scams, and after the Browser paste scam), same severities — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 157 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - Docs in the same push: admin guide gains the 2.5.229 card (the 2.5.227 card retained as history); homepage "New in" rotation (this section, 2.5.227, and the Linux 1.32.0 + Android 1.40.0 client release visible; Server 2.5.226 heads the fold and the fold summary names it); the Spanish homepage gains the 2.5.229 summary paragraph at the top of its Novedades list.
Version 2.5.228 (admin-only)
The trusted-signer catch-up — 20 more verified vendors, 38 → 58 names. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged, and the homepage "New in" headline still tracks 2.5.227.)
- Why a catch-up: the daily expansion's runs of Oct 26–29 were all blocked — first by the stale local tree, then by the concurrency guard doing its job while other releases were in flight. Austin authorized adding the four missed days at once, at the standing rate of 5 verified names per day: 20 names in this one release, no more.
- 20 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: Agilebits (1Password MSI, downloads.1password.com; cert SHA-1 E740F0286402B77268C1D54C0F39AB857F0FABDD), Autodesk, Inc. (Fusion Client Downloader, dl.appstreaming.autodesk.com; cert SHA-1 4A9E6CD4B967ACFF54EC7D3213C061F925682B54), Bitwarden Inc. (Bitwarden desktop installer, vault.bitwarden.com; cert SHA-1 80375A0C9630A51ECB7EC79B37A8174C8DACCCED), Blizzard Entertainment, Inc. (Battle.net Setup, downloader.battle.net; cert SHA-1 A5D44A482C16A2EFA3F025E43C5693E9F47B9E12), Docker Inc (Docker Desktop Installer, desktop.docker.com; cert SHA-1 B6BD29272B07AD4D0F1322A739499D67CA3BAC3F), DOTPDN LLC (paint.net installer, paintdotnet release channel; cert SHA-1 289E2A7D269BAC65754AC274712084385639303C), Electronic Arts, Inc. (EA app installer, ea.com; cert SHA-1 E57750ADD3C6687D4E52466A23DFE2A3C2A22BF8), GOG sp. z o.o (GOG Galaxy, webinstallers.gog-statics.com; cert SHA-1 DF1CBCF2B1AF036C55D122A88A2719D3227916A8), Kovid Goyal (Calibre MSI, download.calibre-ebook.com; cert SHA-1 980637D5ACC6319BEAE773CAC5291EDC3E84DC8E), LogMeIn, Inc. (GoToMeeting Opener, meet.goto.com; cert SHA-1 8F085C938714DD65A4F6C634F22D99D79DC0C7E6), OBS Project, LLC (OBS Studio installer, cdn-fastly.obsproject.com; cert SHA-1 178EB178111547B52B3821A278653EEDC3E5B1F7), Plex, Inc. (Plex desktop, plex.tv; cert SHA-1 C07923815FDA04BEFE83E09CC75E2E936A831726), RealVNC Ltd (RealVNC Connect MSI, downloads.realvnc.com; cert SHA-1 787E284FC93F1F03001F0F9F54467D580B7B7B57), RingCentral, Inc. (RingCentral app, app.ringcentral.com; cert SHA-1 EF1419397607E334403B5943B9398184C9469C42), Riot Games, Inc. (Riot Client, lol.secure.dyn.riotcdn.net; cert SHA-1 DCC7EC946ECDF44D122B146977156BB363BFCEB4), Roblox Corporation (RobloxPlayerLauncher, roblox.com; cert SHA-1 813CA29445456DC3447C173347A0CE5B9494B24C), Signal Messenger, LLC (Signal Desktop, updates.signal.org; cert SHA-1 8A5A56EFFDC462AE8A6CF732BB21E2541995BF36), Slack Technologies, LLC (SlackSetup, slack.com; cert SHA-1 0F8E24482418F503354682F36FD6298BFFE6C677), Splashtop Inc. (Splashtop Streamer, download.splashtop.com; cert SHA-1 D458B32F6946DBB682A9687076EB9209979BDF76), and nordvpn s.a. (NordVPNSetup, downloads.nordcdn.com; cert SHA-1 6C8460C71705A019F5F78AF957A57E55D564FB3A). The full evidence trail (artifact hashes, sources, certs) is in the build ledger for this run.
- The spellings are the certificates', not ours: GOG's cert spells "GOG sp. z o.o" with TWO spaces and a lowercase "sp" — independently re-verified at the codepoint level during this build, because a tidied "GOG Sp. z o.o." would match nothing. Docker Inc carries no trailing period where Bitwarden Inc. and Splashtop Inc. carry theirs; "nordvpn s.a." is lowercase as certified; "Agilebits" is a single word; "DOTPDN LLC" is all caps. Three of the 2.5.211 run's failed attempts landed this time: Signal (its distribution moved to versioned builds on updates.signal.org), OBS Studio (obsproject.com's own CDN link), and 1Password (its Windows artifact is an MSI now).
- Existing installs gain exactly the twenty names once, via a new one-time meta gate (
trustedsigners_expanded_25228) that INSERT OR IGNOREs just the twenty names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Verified but deliberately NOT shipped — the cap is the cap: 6 more names passed the same verification in the same pass and are banked in the ledger for the next daily runs: ELEMENT CREATIONS LIMITED, Figma, Inc., Postman, Inc., Dynalist Inc (Obsidian), Sublime HQ Pty Ltd, and Wire Swiss GmbH. 26 verified, 20 shipped, 0 padded.
- Attempted and deliberately NOT added this run: WhatsApp (Microsoft-Store-only now — the only artifact its download page offers is a Microsoft-signed Store stub, so no WhatsApp signer exists to add), Kodi (its official Windows installer is unsigned), Amazon Kindle for PC (Amazon's download page bot-walls the build's fetcher and no vendor binary could be obtained), and Foxit PDF Reader / Proton VPN (their file hosts return empty replies through the build's network route — both stay candidates for a future run). A candidate with no vendor-served signed binary does not ship.
- Docs in the same push: admin guide gains the 2.5.228 card above the 2.5.211 card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.227, and the fold still heads with 2.5.225).
Version 2.5.227
The Threat Guide's daily update — 3 genuinely new entries, 151 → 154, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.32.0, and Android 1.40.0 are unchanged.)
- Card-tap phone scam (NFC relay) (Medium, Money-driven threats — Dutch police, October 2026; Erste Bank customer warning, October 16, 2026): an automated, in the Dutch cases AI-generated, bank voice warns of suspicious activity and asks the victim to press 1; a live fake employee then says the card is blocked, has the victim download an app, and has them hold their own contactless card against the back of their phone to "verify" it. The app relays the card's live NFC transaction data to an accomplice at a cash machine, who withdraws cash while the victim is still on the call, card still in hand. Dutch police say one ring took €380,000 from 170 victims; a variant opens with a fake prize instead of a fraud warning. Distinct from the relay-attacks entry (a crew relays a card or key fob near the victim, without the victim's help or knowledge) and from mobile-wallet enrollment (a shared bank code enrolls the card in the scammer's wallet): here the victim is talked into performing the relay themselves, with a malicious app as the reader.
- Card courier scam (cut-card / chip pickup) (Medium, Money-driven threats — banks' October 2026 warnings, M&T Bank among them; the FBI courier alert relayed by Universal 1 Credit Union, May 27, 2026, and AARP): a caller who appears to be the bank's fraud department — spoofed number, real account details in hand — says the card is compromised, tells the victim to cut it up but leave the chip intact, asks for the PIN to "cancel" it, and sends a courier or "bank employee" who arrives within minutes, sometimes cutting the card at the door and keeping the chip. Chip plus PIN is the account, usable at a cash machine at once or months later. Distinct from gold-bar courier scams (what is handed over there is savings converted to gold) and from card skimmers & shimmers (a device copies the card unseen; here the victim hands the working chip over in person, reassured by the cutting).
- Browser paste scam (crypto swap hijack) (High, Tricks, traps & nuisances — Cisco Talos, September 2026; BleepingComputer's earlier SwapZone reporting): a "leaked" guide promises an arbitrage flaw — roughly 38% higher payouts, or a loyalty bonus — on a genuine crypto swap site (Talos tracked guides naming SwapZone and SimpleSwap). The victim opens the real exchange, types
javascript: into the address bar, pastes the guide's code and presses Enter, or installs the code as a Tampermonkey userscript. The code runs with the page's own trust: it swaps the deposit address the exchange shows and copies, fakes the rates so the "exploit" looks like it is working, and in the later version Talos documented it fetched fresh payloads from a public Google Sheet (the Google Visualization API as a payload channel). Nothing is installed on the computer and no fake site is involved — the hijack lives only inside the real page in the victim's own browser, and bitcoin sent to the swapped address cannot be recalled. Distinct from ClickFix, FileFix, CrashFix and InstallFix, whose pastes run commands on the operating system: this paste runs only in the page, and its target is a single live transaction. - Danger totals now read Critical 12 / High 59 / Medium 69 / Low 14. The two card entries rate Medium, in line with the guide's other card and courier scams (card skimmers, gold-bar courier scams, mobile-wallet enrollment — all Medium): the loss can be an account emptied, but the defense is a single absolute rule a victim can learn. The browser paste entry rates High, with the rest of the paste family (ClickFix, FileFix, CrashFix, InstallFix — all High): the victim is made to execute code, and the theft completes in the same session.
- Checked and NOT duplicated (2026-10-29 round): PHP2Shell, the Linux Fragnesia chain and Node2Shell / Brotli (CISA KEV additions, October 26–28 — already inside exploit, following the 2.5.208/2.5.226 precedent), the MadApe Microsoft device-code phishing kit (Group-IB, October 29 — already inside device-code-phishing), the Social Security "your statement is ready" email surge (SSA Office of the Inspector General, October 2026 — already inside phishing and gov-impersonation), the Veterans Savings Program enrollment pitch (already inside gov-impersonation and Medicare & health-benefits scams), and the Cisco ASA / FTD SSL VPN zero-day (already inside exploit). 154 is the honest count.
- Coverage claims stay honest: the card-tap entry claims only the Android app scan (signature and behavior, at install and on demand) and scam-call warnings; the card-courier entry claims scam-call warnings and the Senior Protection Pack's trusted-contact alerts, and says plainly that no antivirus can take back a handed-over chip; the browser paste entry claims Clipboard Guard (a copied address changed behind your back), web protection against known lure domains, and the browser extension auditor, and says plainly that code pasted into a page runs as that page and no scanner can make it safe. Each entry gives the rule that defeats it outright: a bank never asks you to tap your card on your phone, never sends a courier for a card and never asks for a PIN, and no guide unlocks a bonus for pasting code into a browser.
- Full EN+ES parity from day one: the Spanish edition gains all three entries in the same release — same slugs, same positions (card-tap immediately after Card skimmers & shimmers, card-courier immediately after Gold-bar courier scams, browser-paste immediately after AI agent instruction hijacking), same severities — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 154 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - Docs in the same push: admin guide gains the 2.5.227 card (the 2.5.226 card retained as history); homepage "New in" rotation (this section, the Linux 1.32.0 + Android 1.40.0 release, and Server 2.5.226 visible; Server 2.5.225 heads the fold and the fold summary names it); the Spanish homepage gains the 2.5.227 summary paragraph at the top of its Novedades list.
Version 2.5.226
The Threat Guide's daily update — 1 genuinely new entry, 150 → 151, in English and Spanish alike. (Server-only release; Windows client 2.5.156, Linux 1.31.1, and Android 1.39.0 are unchanged.)
- 1 new entry, in the house format (what it is / how it gets in / how AEAntivirus stops it): AI agent instruction hijacking (High, Tricks, traps & nuisances — Forever Security, October 22, 2026; The Hacker News, October 26; BleepingComputer, October 25): coding agents (Codex, Claude Code, Cursor and the like) read instruction files — AGENTS.md, CLAUDE.md — in the folder they start in and follow them as orders. A planted instruction file, dressed up as ordinary engineering notes, can order the agent to install a malicious package, rewrite the agent's own machine-wide settings so the attacker's server loads in every later session, or redirect the agent's traffic and expose its login token. A web page can save a file into your Downloads folder on its own, without a click or a permission prompt — researchers proved that route with a harmless test file — and the attack needs one more step from you: launching an AI agent with Downloads (or any folder holding a planted file) as its working folder. Researchers caught a real case in October 2026: an AGENTS.md file in a customer's Downloads folder ordered an agent to install a known-malicious npm package (function-color, from the MALFEX campaign), and in lab tests the trick worked against Codex about 9 times in 10. Distinct from InstallFix (a fake install page whose Copy button swaps the command the victim pastes) and from AI chatbot impersonation scams (the chatbot as the scammer's storefront): here the tricked party is the agent itself, obeying a planted instruction file, with persistence via the agent's own machine-wide config.
- Danger totals now read Critical 12 / High 58 / Medium 67 / Low 14. The new entry rates High, not Critical: it hands an attacker code execution and persistence through a tool the victim trusts, but it still needs the victim to launch an agent from the poisoned folder — one human step stands between the planted file and the compromise.
- Held one day, on purpose: this entry was drafted in the 2026-10-27 daily run and NOT shipped that day, because the local working tree was stale at server 2.5.212 while GitHub main was at 2.5.225 — shipping from the stale tree would have regressed the version and overwritten 2.5.213–2.5.225. The tree was synced to main that evening (byte-verified), so the entry ships now as 2.5.226, exactly one patch past main.
- Checked and NOT duplicated (2026-10-28 round): Node2Shell / Brotli (CVE-2026-111497, CISA KEV October 27) and Adobe Commerce "PolyShell" (CVE-2026-48356, CISA KEV October 27) (already inside exploit), the Convertr analytics-script checkout skimmer (BleepingComputer, October 27 — already inside web-skimmer), fake US District Court ADR notices delivering LodaRAT to 100,000+ businesses (ThreatDown, October 28 — already inside gov-impersonation and spear-phishing; fake courts were already named in gov-impersonation and jury-duty threats were already rejected in 2.5.202), the FTC's GovAssist passport-service impersonation via paid search ads (FTC, October 22 — already inside seo-poisoning and gov-impersonation), the Miasma npm worm that inserts hooks into AI coding assistants (JFrog — already inside malicious-packages, supply-chain, and worm), the Gheist campaign printing GitHub Actions secrets into logs (already inside supply-chain), Skyw4lker device-code phishing (already inside device-code-phishing), fake Firefox crypto-wallet extensions (already inside malicious-extension and wallet-stealer), and the email-bomb + Microsoft Teams IT-support combo (already inside email-bombing and vishing / tech-support-scam). 151 is the honest count.
- Coverage claims stay honest: the entry claims only shipped layers — the signature database naming the known-malicious package, behavioral scoring, API-sequence analysis, and script detection judging the payload the install runs (the same fileless defense that covers the ClickFix family). It also says plainly what no software counter can do: no antivirus can stop an agent from reading a text file, and the instruction file itself is not malware. The advice that defeats it outright is a habit, and the entry gives it: never launch an AI agent from your Downloads folder — work only inside a narrow project folder you created — read any instruction file you did not write yourself before the agent runs, and keep the agent's approval prompts and sandbox protections switched on.
- Full EN+ES parity from day one: the Spanish edition gains the entry in the same release — same slug, same position (immediately after InstallFix), same High severity — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 151 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - Docs in the same push: admin guide gains the 2.5.226 card (the 2.5.210 card retained as history); homepage "New in" rotation (this section, 2.5.225, and 2.5.224 visible; Server 2.5.223 heads the fold and the fold summary names it); the Spanish homepage gains the 2.5.226 summary paragraph at the top of its Novedades list.
Version 2.5.225
Stale pages fixed — the site now always serves its current version, and the holiday decoration is back to its original size (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged).
- What was actually wrong. The 2.5.222/2.5.224 holiday decoration painted correctly all along — it was proven in a real headless-Chromium render of the shipped page (thousands of themed pixels in the garland band and hero). What Austin's normal browser tab showed was a *stale cached homepage*: incognito (no stored copy) and his phone (fresh profile) rendered the live site correctly. Two mechanisms compounded: (1) no response from this server carried a
Cache-Control header — rendered HTML went out with only an ETag, leaving browsers free to reuse pages heuristically; and (2) the site's service worker (registered on every public page since 2.5.134) served / and six other public pages cache-first — its fetch handler returned the cached page with no network check at all, and /sw.js itself carried no cache header, so even the versioned-cache rescue could arrive late. - The fix. Rendered HTML now defaults to
Cache-Control: no-cache, must-revalidate on every extensionless GET/HEAD route outside /api/* and /updates/* (customer, admin and reseller pages alike; handlers that set their own cache header still win, and static assets keep their caching). The service worker's pages are now network-first — the fresh page is served and refreshes the cache, the cached copy is only an offline fallback — while static assets stay cache-first under the versioned cache name, and /sw.js itself is served no-cache. The client update manifests (/updates/manifest.json and the per-platform manifests) also answer no-cache, so an update check can never be served a stale manifest. - Decoration restored to the original scale. At Austin's direction, the 2.5.224 prominence blow-up is reverted wholesale: the theme system is byte-for-byte the 2.5.223 state — garland 44px back to a 16px strip (12px on mobile), 16 motes back to 12 at the original 22–38px sizes and .40–.55 peak opacities, background-pattern opacity back to .45–.55 with the 190px drift tile, card marks back to 24px @ .45 (feature tiles 20px), divider motif 22px, footer row 18px @ .5, and the 64px hero cluster 2.5.224 added is removed (the seasonal line keeps its original 22px flanking ornaments). The 2.5.223 rule — floating 0/1 digits suppressed while a theme is active — is unchanged. With fresh-page delivery, the original decoration is the experiment: a normal tab now shows the current page.
Version 2.5.224
The holiday decoration gets properly visible (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged).
- Everything the themes do, scaled up two to three times. The 2.5.221/2.5.222 decoration was all there but read as invisible in a real photo of the site: the garland strip was 16px tall with ornaments under 10px, the sixteen drifting motifs peaked at 40–55% opacity inside a decoration layer itself dimmed to 45–55%, and the card-corner marks were 20–24px at 45%. The garland is now a 44px-deep swag, the motes are 44–64px at 80–95% peak opacity (sixteen of them), the background pattern layer runs at 80–85% with a larger motif tile, card and feature-tile corner marks are 42px/36px at 85%, divider chips carry a 38px motif, the footer motif row is 30px at 85%, and a pair of the holiday's marks (64px) now flanks the homepage's seasonal line like a design element.
- Same rules, re-pinned with minimums. It is still one switch (live deal + Theme toggle), still decoration only — the button standard's CSS is byte-identical themed vs unthemed, body text keeps the standard colors, nothing is emitted at all when no theme is active, admin and reseller surfaces stay bare, reduced-motion stills the ambient layer, and small screens keep a visible but smaller set. The new v25224 suite pins the prominence minimums (garland height, mote size/opacity/count, pattern opacity, card-mark size, hero cluster) so a future change cannot quietly shrink the decoration back to invisible.
Version 2.5.223
The floating zeros and ones sit out the holidays (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged).
- While a holiday theme is active, the homepage's binary-bits spawner is not emitted at all. The homepage's base ambient effect — 0s and 1s drifting across the page, popping the floating viruses on contact — is gated server-side on the same
activeHolidayTheme() determination that drives the theme itself (applicable deal live + its Theme toggle on). Themed pages carry no spawner markup at all; the theme's own motes drift alone. - The gate is "theme active", not "deal live". A live deal with its theme switched off keeps the digits, and so does every no-deal day — the unthemed homepage emits the spawner byte-for-byte as before. The floating viruses themselves stay in both states; only the digits sit out.
- The digits only ever lived on the homepage (the English and Spanish views) — no other page carries the spawner, and nothing about the admin panel, reseller pages, prices or buttons changes.
Version 2.5.222
More holiday decoration, across the whole site (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged).
- Every theme pack gains a full decoration system. 2.5.221's accents and corner ornaments grow into: a themed garland under the header on every storefront page (pumpkin-and-spider trim at Halloween, a string of lights at Christmas, flag bunting for Memorial Day and the Fourth, hanging price tags on Black Friday, stars for the New Year, pencils and books for Back to School), an ambient layer of the holiday's own motifs drifting gently down every page (twelve per page; transform/opacity animation only; behind all content; pointer-events-none), a centered motif ornament on section dividers and feature-group tops, motif corners on cards and feature tiles, themed marks at the ends of the trust strip and the catalog nav, a decorated footer rule with a motif row, and ornaments flanking the homepage's seasonal line.
- One coherent motif set per holiday. Each pack now defines three small inline-SVG motifs in its accent palette — a *mark* (reused by the dividers, card corners, trust strip, footer row, and seasonal line), a *garland* tile, and a *drift* tile for the ambient layer — so every decoration on the page speaks the same visual language. No external images anywhere; the whole theme stylesheet is still emitted only while a theme is active, and out of season the pages remain byte-clean of it.
- Same guardrails, re-pinned. The button standard's CSS stays byte-identical themed vs unthemed and the theme block still contains no
.btn rule; decorations frame the content and never sit over text or buttons; prefers-reduced-motion now stills the ambient layer entirely; on small screens the garland shrinks and half the drifting motifs hide. Admin and reseller surfaces remain fully excluded, the Theme toggle still governs everything (off = plain banner + discount, no decoration at all), and the Terms are unchanged — presentation only. - Docs in the same push: homepage "New in" rotation (this section, 2.5.221, and 2.5.220 visible; 2.5.219 heads the fold) in EN+ES, and an added sentence in the Help & FAQ holiday-sales answers (EN+ES). The admin guide is unchanged — the Theme toggle's meaning did not change.
Version 2.5.221
Holiday themes: while a holiday sale is live, the site dresses for the holiday (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged).
- One switch drives discount, banner, and theme. A theme is active exactly when its deal is the live deal the site may promise (live by the same calendar resolution as 2.5.219, Stripe coupon known to exist) AND the deal's new Theme toggle is on. Disabling a deal in the admin panel kills its discount, banner, and theme together; switching Theme off leaves the discount and the plain (unthemed) banner running. The toggle sits next to the deal's enable/percent/window controls in Customers → Holiday deals, persists in
DATA_DIR/holiday-deals.json, and takes effect with no redeploy. - Seven theme packs, data-driven. Halloween, Black Friday, Christmas, New Year, Memorial Day, Independence Day, and Back to School each carry an accent palette (CSS variables for the header rule, hero eyebrow, trust strip, catalog nav, and footer rule), a restrained CSS/SVG-only decoration (falling snow at Christmas, twinkling stars for New Year, corner ornaments and faint patterns elsewhere — no external images, no layout shift, pointer-events-none, reduced-motion respected), a themed banner line, and a seasonal homepage line — all in English and Spanish. The banner keeps the deal's facts and the honesty line ("First payment only — renews at the regular price") and gains one holiday clause: "No tricks, just a treat" (Halloween), "Our biggest markdown of the year" (Black Friday), "A little gift from our family to yours" (Christmas), "Start the year protected" (New Year), "With gratitude, as we remember" (Memorial Day), "Celebrate the Fourth, covered" (Independence Day), "Back to class, back to covered" (Back to School).
- Guardrails. Themes are additive decoration only: they define their own
--th-* variables and contain no .btn rule, so the standard site button (bright cyan, dark text, rounded, glow) is never restyled, and body text keeps the standard colors. No theme markup or CSS is emitted at all when no theme is active — zero residue out of season. Admin pages (/admin, /admin-login, sub-pages) and reseller surfaces (custom domains and the /reseller portal) never receive a theme. - Docs in the same push: homepage "New in" rotation (this section, 2.5.220, and 2.5.219 visible; 2.5.218 heads the fold) in EN+ES, a theme line in the Help & FAQ holiday-sales answers (EN+ES), the admin guide's Holiday deals paragraph, and the admin section's own intro. Terms unchanged — themes are presentation only.
- Heads-up for deployment: production is on 2.5.219, so one redeploy brings 2.5.220 and 2.5.221 together; landing before Oct 31 turns the Halloween theme on for the tail of its window.
Version 2.5.220
One threat-feed count, everywhere (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged — existing clients pick the corrected number up on their next threat-intel check, no client update needed).
- The client and the site were counting different things. The app's "Threat intel: N feeds" line came from
/api/threatintel-status, whose summary counted *every* source row in the database — auto-paused feeds included — while the homepage counted "feeds actually on" (sources not auto-paused, plus the admin intel feed when configured, plus the MalwareBazaar import when its key is set). Same server, two numbers (17 in the app vs 18 on the site). The endpoint now takes its summary count from the same shared helper the homepage uses (liveFeedCount()), so every surface publishes one number. The endpoint's per-feed list still shows every source row, and its paused count still counts auto-paused rows — only the headline summary was wrong. - A dead bonus, revived. Inside that helper, the admin-intel-feed +1 had silently never fired: it read a file path constant that was never defined, inside a try/catch that swallowed the error. "Configured" now means what the rest of the server means by it (the feed-health report's test — an intel feed URL is set), so the shared count matches its documented definition on both surfaces.
- No more typed fallback number. When the live count can't be computed, the homepage trust strip used to print a hardcoded "16" — a made-up figure that could silently disagree with everything else. Following the signature-version figure next to it, the strip now simply omits the feed figure when the count is unavailable (same on the Threat Intelligence card, EN+ES) rather than inventing one.
- The Spanish homepage's Threat Intelligence card was showing raw code. Its count expression sat inside a plain quoted card string, where
${...} never evaluates — the live ES page literally printed ${feedCount != null ? feedCount : 16} fuentes por defecto… instead of a number (verified on production during this release). The card now takes the count through the view's placeholder substitution (%%FEEDS%%, alongside the existing %%B%% brand placeholder), so it renders the real shared count, or just "Fuentes" when the count is unavailable. A comment in the view and a test pin guard the pattern. - Admin label clarified. The Feed sources (MISP/STIX) header intentionally shows the raw source-row count for the table under it; it now says "source rows" so it can't be mistaken for the published feeds-on number, and its auto-paused breakout (and the per-row paused badge) actually work — the comparison checked a state value the data never holds, so they silently never appeared.
- The holiday deal banner no longer shows on the admin panel. The 2.5.219 banner was injected by the shared page wrapper, so it appeared on staff pages too. The deal banner is now excluded on admin paths (
/admin, /admin-login, and any admin sub-page) and stays exactly as before on customer-facing pages. Only the deal banner is scoped: the separate announcement banner and the impersonation safety banner are untouched. - Docs in the same push: homepage "New in" rotation (this section, 2.5.219, and 2.5.218 visible; 2.5.217 heads the fold) in EN+ES. Help/FAQ state no feed count or counting rule, so they needed no change.
Version 2.5.219
Premade holiday sales that switch themselves on and off by date (server-only release; Windows client 2.5.156, Linux 1.31.1 and Android 1.39.0 are unchanged).
- Seven premade deals, fully automatic. Halloween (Oct 25–31, 15% off), Black Friday (the Friday after Thanksgiving through Cyber Monday, 30%), Christmas (Dec 20–26, 25%), New Year (Dec 29 – Jan 2, 20%), Memorial Day (the Saturday before Memorial Day through Memorial Day, 15%), Independence Day (Jul 1–5, 20%), and Back to School (Aug 15 – Sep 5, 15%). A deal is live purely from the calendar — America/New_York, whole days inclusive at both ends, resolved at request time. No cron, no scheduled job, no manual activation. Floating holidays are computed (Thanksgiving = 4th Thursday of November; Memorial Day = last Monday of May). Overlapping windows never stack: the single highest percent wins, ties go to the earliest start date.
- First payment only — renewals stay honest by construction. A live deal takes its percent off the FIRST payment of a NEW plan subscription (the standard /subscribe plan checkout), applied as a Stripe coupon with duration
once, so every renewal bills the regular price and the flat-renewal promise ("your price is the price that renews") stays literally true. Add-on checkouts, gift purchases, renewals, and reseller flows are never discounted. A live deal REPLACES any other discount for that checkout (promo code, referral free month, Custom volume tier) rather than combining with it — one discount per checkout, as Stripe allows. - Site presentation. While a deal is live, a site-wide banner (EN+ES) names the deal, the percent, and the end date, with the line "First payment only — renews at the regular price," and /subscribe shows each plan's regular price struck through next to its deal first-payment price, computed from the real plan prices (the tier table / per-device rate — no hardcoded amounts). When no deal is live there is no banner and no residue: regular prices, exactly as before.
- Fails closed, never loud. Each deal's Stripe coupon is created lazily under a fixed id (per deal + percent), retrieved-or-created idempotently, and the id is stored in the persisted deal config. If Stripe is unreachable when the coupon would be created, the DISCOUNT fails — not the checkout: the customer buys at the regular price, the banner stays down (the site never promises what checkout can't do), and a later request retries creation.
- Admin override (Customers → Holiday deals). All seven deals listed with their windows, percents, computed dates for the current cycle, and LIVE NOW / next-upcoming status, plus a preview-as-of date readout. The owner can disable a deal (disabling the live one ends it on the site immediately), edit its percent, or edit its window (fixed dates as MM-DD; floating windows as day offsets from the holiday). Edits persist in
DATA_DIR/holiday-deals.json (the policy.json pattern — survives redeploys) and take effect with no redeploy. - Docs in the same push: homepage "New in" rotation (this section, 2.5.218, and 2.5.217 visible; 2.5.216 heads the fold) in EN+ES, Help & FAQ sales copy in both languages (including the renewal answer, which now names holiday sales as the first-payment-only exception), a Terms (§3) holiday-sales paragraph, and this admin guide entry (Customers → Holiday deals).
- Heads-up for deployment: the Halloween window is Oct 25–31 — redeploying during that window turns the Halloween deal on the moment the new server answers.
Version 2.5.218
_Gifted months and the free trial (server) + a Windows client tidy-up (Client 2.5.156)._
- A gifted month closes the Free plan's trial. If any license on an account was gifted time — extended by an admin without payment, or bought as a gift and redeemed into the account — the account is no longer eligible for the Free plan's 7-day full trial: they already got a free month, so there is no free trial on top. The rule is account-wide and permanent (a fresh Free key on the same account doesn't reopen it), every surface agrees (the app banner is not offered, My Account shows why instead of the start button, and both start actions refuse with their own message — "This account received a gifted month, so the free trial isn't available."), and past admin extensions recorded in the audit log were marked when this shipped. The Router Guard 30-day trial is a different trial and is not affected.
- Windows client 2.5.156 — Status page layout. The live "Skipped as unchanged" scan counter no longer crowds the Status page buttons: the live scan lines now sit in their own reserved block above the button row with clear space. Button labels always fit their buttons now — long labels step down in size or wrap instead of being cut off with "…", and the Smart Scan button shows its own name again (a status line was being written onto it by mistake).
Version 2.5.217
Router Guard shortlist — the audit's 11th universal check (DNS rebinding), a third Netgear deep check (WPS), and an IPv6 outside-in WAN test. (Windows client 2.5.155; Linux client 1.31.1; Android 1.39.0 is unchanged.)
- DNS rebinding check (universal check #11 — Windows and Linux): the audit now asks the GATEWAY's own DNS resolver for the test name
rebind.aeantivirus.com. An answer carrying a private, loopback, or link-local address is a Warn (the resolver hands out inside addresses for public names — exactly what DNS rebinding attacks rely on to reach devices inside the network from a website); an empty or 0.0.0.0 answer is a Pass; NXDOMAIN, a timeout, or a public-only answer is N/A. Until the test name exists in DNS the check reports N/A on every audit — it is present and waiting, and the copy says exactly that; it never claims live protection from a check that cannot run yet. Austin's step: create the DNS record — host rebind.aeantivirus.com, type A, value 192.168.0.1 — and the check starts measuring on the next audits. - Netgear WPS deep check (Windows client 2.5.155, third adapter check — 14 total on Netgear/Windows hosts): on Netgear routers with the sign-in saved on a Windows main, WPS joins guest network and remote management: best off, or on only with the PIN method disabled; an unanswered question is N/A with the score renormalized, never a guessed grade. Overlap audit (already established): NO overlap with the existing Wi-Fi checks — there were zero WPS hits in
src/ outside web-protection IDs, wifisec grades only the connection the PC has joined, and the Tech Tools Wi-Fi radio check emits local radio fields only; WPS state lives on the router and only the Netgear adapter can ask it. Linux mains run the 11 universal checks (the deep checks stay Windows-only). - IPv6 outside-in WAN test (server): the WAN test gains an IPv6 twin ("Test what the internet sees (IPv6)") in My Account, the client routes, and the admin panel. Its targets are hard-bound to the public IPv6 addresses the network's own main device reported in its audit (
ipv6_globals) — no request input can influence the target, any address-shaped field in a request body is ignored, and if the address our server sees the device connect from is IPv6 and is NOT among the reported addresses, the test explains (mismatch) instead of scanning; no reported addresses explains as no-ipv6. It probes the same port set as IPv4 against every reported address (max 4) and cross-checks open ports against the router's own report the same way. Shared rate limit: one WAN test run per account per hour TOTAL across IPv4 + IPv6 (one table, rg_wan_tests, with a new family column; a v4 run blocks a v6 run within the hour and vice versa). Capability honesty: the server self-tests whether it can originate an IPv6 TCP connection (two well-known anycast resolvers, ~2.5s, verdict cached an hour, AE_IPV6_CAPABLE=1/0 overrides) — where it cannot (as production may not), every surface shows "IPv6 test: not available from our server yet" as text, no run row is created, and no result is ever fabricated. Results feed the "What changed" timeline labelled "WAN test (IPv6)". - Server plumbing:
sanitizeRouterAudit now preserves ipv6_globals — each entry re-validated as a syntactically valid, globally-routable IPv6 address under the same rule the clients use (2000::/3 minus link-local/unique-local/loopback and every IPv4-mapped form), deduped, capped at 8; the Free summary reduction still never carries it. Publishing: Windows 2.5.155 (already-built binaries + regenerated delta patches) and Linux 1.31.1 (rebuilt daemon, repackaged tarballs) are in server/updates/ + server/installer/; all manifests keep signatures v108 / sha256 96d9ecb7… untouched. - Docs in the same push: homepage "New in" rotation (this section, 2.5.216, and 2.5.215 visible; 2.5.214 heads the fold) plus the Everything-included Router Guard card (eleven universal checks, three Netgear checks, IPv6 WAN test), Help & FAQ in both languages (rebinding in the checks list, three Netgear checks, the IPv6 test), Terms (rebinding + WPS in the check enumeration, the IPv6 test beside the WAN test), Privacy (the audit's reported public IPv6 addresses are disclosed as stored with the audit), the admin guide's Router Guard card (rebinding incl. the pending-DNS-record state, WPS, IPv6 test), and the client's What's New page (shipped with the 2.5.155 client).
Version 2.5.216
Router Guard now includes DDoS Protection (server-only release; Windows client 2.5.154, Linux client 1.31.0 and Android 1.39.0 are unchanged — no client What's New entry, per the server/client version rule).
- The bundle: any account with Router Guard entitled — an active Router Guard subscription (including its 24-hour payment-failure grace), a running Router Guard free trial, or an admin Router Guard grant — gets DDoS Protection included at no extra charge. The inclusion is computed at the existing entitlement choke points (
addonActive() / accountAddons() in server/server.js) from Router Guard's own entitlement; NO paid_addons row is created for it, so DDoS purchase counts, reseller commissions and revenue reporting never see bundled accounts. The heartbeat carries ddos in the add-on list with ddos_source: 'routerguard-bundle' (vs 'own'), and /api/account/addons exposes the same source, so a bundled activation is never mistaken for a purchase. - Lifecycle follows Router Guard: trial on → DDoS on; RG payment fails → DDoS stays on through RG's same 24-hour grace (at 23h on, at 24h+ off); RG ends (trial over, post-grace pause, cancel, revoke) → bundled DDoS turns off on the same heartbeat — unless the account independently holds DDoS (own subscription incl. its own grace, or admin grant), in which case DDoS simply continues under its own entitlement. The Windows client needed no change: its DDoS engine gates on
cloud_addon_active("ddos") from the heartbeat add-on list and re-checks whenever that list changes (the 2.5.215 generation-counter fix), so bundled inclusion arrives via heartbeat automatically. - My Account DDoS card — kept, price kept: the card keeps its slot and keeps showing its standalone price. With Router Guard it shows an "Included with Router Guard" badge (EN+ES) and the Subscribe button is replaced by a disabled included state — and the checkout route itself refuses with "already included with your Router Guard," so there is no way to accidentally buy what the account already has. An account holding BOTH sees a note that it can cancel the separate DDoS subscription since Router Guard includes it — informational only; nothing is ever auto-canceled. Without Router Guard the card is unchanged except one added line: "Or get it included with Router Guard." Card visibility is unchanged: it renders only under its existing condition (active or
STRIPE_DDOS_PRICE_ID set); with the price ID unset and no own subscription, an RG account's card stays hidden exactly as before — the bundled entitlement itself works regardless, because it rides the heartbeat, not the card. - Copy: the Router Guard card, its catalog blurb and Terms text, the homepage Everything-included Router Guard card and DDoS paragraph, the Help & FAQ Router Guard answer (EN+ES), the Terms Router Guard paragraph, and the admin guide's Router Guard card each gain the plain line: Router Guard includes DDoS Protection. No price changes; the catalog is unchanged at 128 paid + 100 free = 228 (an entitlement bundle, not a merge).
- Tests: new suite
v25216-rg-ddos-bundle-test.js pins every bundle transition, the card's four states via real /account requests (EN+ES), checkout refusal, Command Center purchase-count purity, catalog counts, the inline-script regression and the served-byte census.
Version 2.5.215
Router Guard phase 3 — the feature is complete: Linux hosts run the audit, Netgear routers get two deeper checks, every network gains a "What changed" timeline — plus one fairness rule for every paid add-on. (Windows client 2.5.154; Linux client 1.31.0; Android 1.39.0 is unchanged.)
- Linux Router Guard host (Linux client 1.31.0): a Linux desktop designated as a network's main device now performs the full Router Guard host duty itself — gateway discovery and fingerprinting (default route, neighbor/ARP table, SSDP description, gateway web banner, MAC vendor), the same scored audit in the same shape the Windows client posts, UPnP mapping enumeration with the router's external address, the DNS/DHCP mismatch logic (a mismatch alone is never a Fail), and the gateway service probe — reusing the network collectors the Linux Tech Tools already ship. It honors the server's per-network main assignment and the add-on entitlement (dormant when unentitled), audits on the same roughly-six-hour cadence with jitter, answers the admin's run-now command, and can be triggered by hand from the daemon's command line. The Netgear adapter ships on Linux in enrichment form: when a router sign-in is saved on the device (a local file the daemon refuses to read unless its permissions are owner-only, set with the daemon's --router-guard-netgear-set command; the password is never logged, never put in any report, and never transmitted to our servers), a Netgear session identifies the router exactly — brand, model, and firmware from the router itself replace the generic fingerprint. The two deep checks below and block/allow stay Windows-only in this release: the Linux audit's scored surface is the universal checks only (eleven since 2.5.217), so a Netgear router behind a Linux main gets the full universal audit with exact router identity, and the guest-network and remote-management checks need a Windows main for now.
- Deeper Netgear checks (Windows client 2.5.154): on Netgear hosts with the router sign-in saved, the audit grows from 10 checks to 12 (2.5.217: 11 universal + the WPS adapter check make 14). Guest network (WLANConfiguration GetGuestAccessEnabled, with the GetGuestAccessEnabled2 variant as fallback): Pass when off, or on with guest isolation reported on; Warn when on with isolation off or unreported (the API exposes no dedicated isolation action — where the firmware doesn't report it, the check says exactly that). Remote management (DeviceConfig GetRemoteMgmtEnable / GetRemoteManagement — no public client library wraps these, so an unsupported answer degrades honestly): Pass off, Fail on, and when the outside-in WAN test has proven reachability the Fail says so with that evidence — the router's own answer alone never claims internet exposure. No sign-in saved, or the firmware doesn't answer, makes either check N/A with the plain reason and the score renormalizes over the checks that did run — never a guessed Fail. The Windows audit now also reports the network's DNS servers, which the timeline below diffs. Universal copy no longer claims a flat "10 checks": Netgear hosts run the universal checks + up to 3 adapter checks (11 universal + 3 = 14 since 2.5.217).
- "What changed" timeline: for entitled accounts the server now keeps a bounded history per network — a snapshot of each host audit (score, per-check states, port-opening identities as protocol + external port + a short one-way hash of the internal device — never the raw address, DNS servers, router brand/model/firmware), capped at the newest 100 snapshots and 90 days per network, pruned in small batches. Change events are derived at ingest against the previous snapshot: score moves, named check flips, a port opening appearing or disappearing, a DNS set change, a router model or firmware change; WAN-test results land as their own events; the change log is capped at the newest 300 per network. One-off manual audits are marked and excluded from the trend — the timeline is the main device's monitoring record, and it is history only, never a second alert channel (the existing findings alerts are unchanged). My Account shows each network's score trend and latest changes in plain language (English and Spanish), the admin Router Guard view carries the same history per network, and the Windows client's Router Guard page shows the latest few changes for its network. Unentitled accounts store the summary only, exactly as before — no timeline. The Privacy Policy discloses the retention in this release.
- One entitlement rule for every paid add-on (Austin, 2026-10-26): a paid add-on runs only when it has been activated — an active subscription, an admin grant, or its own free trial running. The audit behind this release found the gap it closes: when an add-on's own subscription payment failed, the subscription lifecycle webhook canceled the add-on on ANY non-active Stripe status — instantly, on the first past_due (and it would also have canceled a subscription sitting in its trial window). Now the first failed payment starts a 24-hour grace clock on that add-on (paid_addons.pay_failed_at; Stripe retries inside the window never reset it): the add-on keeps running for 24 hours from the first failure, then drops out of the account's add-on list — and therefore out of the client heartbeat — until the payment clears, which reactivates it automatically and clears the clock. Router Guard's entitlement folds into this same rule; it has no separate payment handling. Grants and one-time purchases have no recurring payment to fail and are untouched; the main-license rules are unchanged (an expired main license still empties the list, the 2.5.195 fallback still keeps independently-paid add-ons whose own payments are healthy, and a pause still suspends everything immediately). Nothing goes dark silently: My Account shows a payment-issue banner and per-card state ("keeps running until …" inside grace, "Paused — payment failed …" after it) with an Update payment button, the admin Command Center lists payment-failed add-ons as their own state, and the admin payment-failed alert now names the add-on and its grace deadline. Client side, every paid feature already follows the heartbeat add-on list — with one exception found and fixed in this release: DDoS Protection had cached its entitlement the first time it was checked and never re-read it (its invalidate hook had no callers), so it could fail to start after purchase and keep running after removal until the app restarted; it now re-checks whenever the heartbeat's add-on list changes.
- Docs in the same push: homepage "New in" rotation (this section, 2.5.214, and 2.5.213 visible; 2.5.210 moves into the fold) plus the Everything included Router Guard card (Linux mains, the two Netgear checks, the What changed history), Help & FAQ entries in both languages (Linux mains, the Netgear checks, the What changed timeline, and the add-on payment grace), Terms updates (add-on payment terms with the 24-hour grace in plain words; Router Guard on Linux mains and the Netgear checks), the Privacy Policy's Router Guard retention disclosure, the client's What's New page, and the admin guide's Router Guard card.
Version 2.5.214
Router Guard grows up: a WAN test that checks what the internet sees, one main device per network, and desktop-only mains. (Windows client 2.5.153; Android 1.39.0 is unchanged; Linux client 1.30.1 reports its form factor the same way — Router Guard for Linux and Android is still queued.)
- Test what the internet sees (WAN test): a new button in My Account (and in the client, and for technicians in the admin panel) asks the server to check, from the outside, which TCP ports on your network's own public internet address the internet can reach — and cross-checks anything it finds against what your router itself reported: a port that matches one of your router's own port openings is named as such, a port matching a service on the router itself is named, and a port your router's report doesn't explain is flagged as worth checking. The test runs only when you press the button, it scans only your own public address — the address our server itself sees your main device connect from, confirmed against the address your router reports; if the two disagree, or your connection sits behind your provider's shared equipment (CGNAT), the test explains that in plain language instead of scanning an address it can't confirm is yours — and every run is logged. One run per account per hour.
- One main device per network: Router Guard's main device is no longer one-per-account — it is one per network. A home and a shop can each have their own main device watching their own router; setting or replacing one network's main never touches the other's. A main chosen under 2.5.213 keeps working: it binds to its device's network the first time that device reports its gateway.
- Desktop-only mains: the main device must be a desktop that stays on the network — laptops and phones can't be set as main (the set-main button explains this instead of offering itself, and the server refuses the request). A device whose form factor isn't known yet can still be chosen (it's flagged as unknown). If a device that is a main later turns out to be a laptop, its main is suspended — not deleted — and My Account and the admin panel prompt for a new main for that network; it resumes if it reports as a desktop again. Windows reports its form factor from the machine's SMBIOS chassis type, Linux from its DMI chassis type — never guessed.
- Run Router Guard now, and offline prompts: technicians can queue a Router Guard audit on demand from the admin panel (it runs on the network's main device at its next check-in), and when a network's main device hasn't been seen for over 48 hours, My Account and the admin panel say Router Guard isn't running on that network and prompt for a new main — a replacement is never picked automatically.
- Unified client button style (Windows client 2.5.153): every button in the Windows app now shares one style — the dark teal-blue fill with a complete cyan outline and a cyan bar on the left edge that the sidebar and Router Guard buttons each had half of; the current page's sidebar button stays distinct (brighter fill, cyan text, and a dot).
- Linux form-factor reporting (Linux client 1.30.1): the Linux client reports its chassis form factor with its heartbeat, the same as Windows, so a Linux desktop can be a network's main device and a Linux laptop is refused the same way.
- Router Guard is now a paid add-on (pricing pass): Router Guard leaves the base plans and becomes a dedicated add-on — $8.33/month, the $99.96/year plan billed monthly (Stripe price via
STRIPE_ROUTERGUARD_PRICE_ID — Austin's step: create that one recurring price and set the env var, then the Subscribe button appears; the trial works without it), with a 30-day free trial once per account, enforced by our server (accounts.rg_trial_used / rg_trial_ends_at — cancelling never re-arms it; there is no reset path). Entitlement is account-level (active routerguard add-on OR a running trial) and drives everything the 2.5.213/2.5.214 wiring gated on plan tier: WAN test access, heartbeat routerguard_mode (full vs summary) plus new rg_entitled / rg_addon_active / rg_trial_days_left fields, and audit storage (unentitled accounts — any plan — store the summary shape only; a Free-plan account inside its RG trial stores full). The one-time audit summary preview and device-local connection tests stay available to every account, and Tech Tools network diagnostics are not gated by the add-on. 2.5.213 was never deployed, so no grandfathering applies. The add-on catalog is now 228 entries: 128 paid + 100 free. - Docs in the same push: homepage "New in" rotation (this section, 2.5.213, and 2.5.210 visible; 2.5.208 moves into the fold), the Everything included Router Guard card, Help & FAQ entries in both languages (per-network mains, desktop-only mains, and the new WAN test Q&A), a Terms update covering per-network/desktop-only mains plus the WAN test disclosure, a new Privacy paragraph for Router Guard and the WAN test, and the admin guide's Router Guard card.
Version 2.5.213
Universal Router Guard — the router itself joins the protection, on any network, with no router login required. (Windows client 2.5.152; Android 1.39.0 is unchanged; Linux client 1.30.0 ships the new Linux network checks — Router Guard for Linux and Android is queued.)
- Any router, no login needed: Router Guard is built into AEAntivirus — part of the paid base plan, not a paid add-on. It finds your default gateway and audits the router itself, whatever the brand — TP-Link, ASUS, Linksys, Netgear, ISP gateways, and mesh systems alike — with no router login required. It fingerprints brand, model, and firmware when they are discoverable, and says "Unknown router" honestly when they are not. The Free plan gets a one-time audit summary only — the score and the names of any failed checks, with no continuous monitoring, alerts, or adapters.
- What it checks: whether the admin page is served over plain HTTP instead of HTTPS; risky services listening on the router (Telnet, FTP, SSH, and remote-management / TR-069); the UPnP state and the full UPnP port-forwarding table, with risky forwards flagged; the DNS servers in use versus the router's own DNS (a possible hijack or mismatch); and signs of a rogue DHCP server or a duplicate gateway.
- A score in plain language, and paid plans keep watching: every audit gives a score out of 100 with a plain-language Pass / Warn / Fail list. On paid plans Router Guard re-checks in the background and alerts you when it sees a new risky port forwarding, a DNS change, a drop in the score, or the gateway changing identity.
- Optional Netgear adapter — your router password never leaves your PC: on Nighthawk routers, if you choose to enter your router admin password, it is stored only on your own PC, encrypted with Windows DPAPI — it is never sent to or stored on AEAntivirus servers. With it, the adapter can show attached devices and block or allow a device by MAC address where the router firmware supports it. Other brands are audit-only for now, and those controls are shown as honestly unavailable rather than faked.
- Only a summary is reported, and the router was the gap: only a summary reaches your account — brand and model, the score, the check results, and the forwarding count — never router credentials, and never LAN inventory beyond what device discovery already reports. Router Guard does not replace the Wi-Fi security check, device discovery, the Network Monitor, or DNS filtering; it covers the router itself, which those features do not.
- One main device per account (Router Guard main-device rule): Router Guard's network-wide background monitoring runs from ONE designated main device per account, chosen by the customer in My Account (badge ★ Main device on that device's card, "Set as main device" on the others, and a banner above the device list while no main is set). The server is the source of truth: the designation lives in the new
account_rg_main table (one row per account, set only by the customer — there is NO auto-pick), and every heartbeat answers is_rg_host (true only for the designated install, plan-independent) plus rg_host_name. POST /api/account/main-device sets/clears it (ownership-checked through installs → licenses → account, audit-logged rg-main-set / rg-main-cleared), and GET /api/router-guard/status (install-auth) lets a device read its own account's host and the host's stored audit. Removing, transferring, license-deleting, merging, admin-deleting, or duplicate-sweeping the main install clears the designation. Safety net: a non-main install that posts a non-manual audit for the SAME gateway as the host is flagged in rg_conflicts and shown as a warning in Admin → Security operations → Router Guard (its audit is still stored as evidence; the client stands down via is_rg_host). A one-time manual check (manual:1 on the audit) can still be run from ANY device and never counts as a conflict. Class split: device-local connection tests can be run on any device at any time — only Router Guard's network-wide monitoring is main-device-only — and Tech Tools network checks (technician-run on the individual device they pick, diagnosing that device's own connection) are exempt from the main-device rule; the diag queue and /api/diag-result paths take no main-device status into account at all. - Docs in the same push: homepage "New in" rotation (this section, 2.5.210, and 2.5.208 visible; Server 2.5.204 heads the fold and the fold summary names it), a Router Guard card in Everything included → Network (English and Spanish), a Help & FAQ entry in both languages, and a Terms disclosure covering the summary, the local-only Netgear credential, and the line between this automated self-check and the paid technician services. The Spanish homepage carries no "New in" archive by design (it links the changelog), so no ES rotation applies.
- Tech Tools network round — 9 new technician network checks: a technician can now also run a ping quality check, a DNS audit (compares the device's resolvers), a DHCP audit, traceroute, the router's UPnP port openings, a gateway service probe, Wi-Fi radio detail, an ARP table audit, and a download speed test — the speed test downloads from the AEAntivirus server only (download only; there is no upload test). Six of the nine also run on Linux (Linux client 1.30.0) (ping quality, DNS audit, DHCP audit, traceroute, ARP table audit, download speed test); UPnP port openings, the gateway service probe, and Wi-Fi radio detail are Windows-only. New catalog totals: 58 checks on Windows, 25 on Linux, 10 on Android (macOS 0). Findings from the new checks dedupe by fingerprint like every other Tech Tools finding, so the same finding does not alert twice — and Router Guard's own local alerts stay separate from Tech Tools findings. The Full checkup preset is unchanged.
Version 2.5.212 (admin-only)
Stale duplicate install-record cleanup — one PC, one fleet record. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged, and the homepage "New in" headline still tracks 2.5.210.)
- Why: registration dedup is scoped to a single license key, so duplicate install rows for the same PC accumulate across keys and from re-registration while an old row is still fresh — the fleet then shows every ghost row (one shop PC appeared as 7 records). A new sweep,
sweepDuplicateInstalls, retires those stale duplicates automatically: once on the first boot of this release (meta gate dupcleanup_boot_25212, set when the run completes, on a delayed unref'd timer — AE_DUP_SWEEP_BOOT_DELAY_MS, default 600000 ms) and then daily via the job journal (dup-install-cleanup). An in-flight flag keeps the boot and daily runs from ever overlapping, installs load in 1000-row pages, and the sweep yields between pages and every 25 retirements. - The rules: records group two ways — an exact-fingerprint group (a normalized device fingerprint, trim + lowercase, shared by 2+ rows) and, over the survivors, a hostname fallback group (same lower-cased hostname + platform, for rows in no fingerprint group). The keeper in each group is the record with the most recent heartbeat (never-heartbeated counts as oldest; ties break by newest creation, then install id). Only a non-keeper whose last heartbeat is missing or older than 24 hours is retired. Never touched: the keeper, any record heartbeating within 24 hours, and the only record of a machine. One safety skip: a ghost that shares the keeper's license key AND fingerprint is skipped (and the skip is audit-logged once per group) — tombstoning it would poison the keeper's own re-registration path under that key and blur restore-by-pair.
- Retired records are tombstones, not deletions: each one goes through the same
revokeInstallRow path as a removal, with reason duplicate-cleanup and an install.duplicate_cleanup audit entry naming the retired id, hostname, keeper, and the rule that fired. They land in Fleet → Removed devices and are restorable there by the owner exactly like any removed device (2.5.186). Because tombstoned rows leave the installs table, their seats free immediately — seat counts are COUNT(*) over installs, so no billing code changes and no Stripe call is made; subscriptions and license rows are never touched. The Boot Guard fleet view now also excludes tombstoned installs, and a restore makes a device's stored Boot Guard report reappear there. The sweep is idempotent: a second run retires 0. - Docs in the same push: admin guide gains the 2.5.212 card (at the top of the guide and beside the 2.5.186 Removed devices note in the Fleet card); the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.210, and the fold still heads with 2.5.203).
Version 2.5.211 (admin-only)
The daily trusted-signer expansion — 5 more verified vendors, 33 → 38 names. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged, and the homepage "New in" headline still tracks 2.5.210.)
- 5 new names, each verified from the vendor's own signed installer — downloaded from the vendor's official channel during this build, with the signer resolved from the file's own signature (SignerInfo) and copied byte-exactly as the certificate spells it, never normalized or guessed: GitHub, Inc. (GitHub Desktop, central.github.com; cert SHA-1 7DABFEF8EB9B3794E45806BC491D8DECF4348270), Logitech Inc (Logitech G Hub, download01.logi.com; cert SHA-1 BFAB6EF922D0DDE848AE747B2AD58506E17D9ED1 — no trailing period, as certified), OpenJS Foundation (Node.js v24.22.1 LTS MSI, nodejs.org; cert SHA-1 6018A47185244F3B175DAFC18D0D8EB47E3F84CF), Telegram FZ-LLC (Telegram Desktop, telegram.org; cert SHA-1 C8CB11E5352916312801039AEDC5F9E8C78E48D8), and win.rar GmbH (WinRAR 7.01, rarlab.com; cert SHA-1 729AE1F8B489DE176CC099FF49937F85F9E412F7 — lowercase "win.rar", as certified). The full evidence trail (artifact hashes, sources, certs) is in the build ledger for this run.
- Existing installs gain exactly the five names once, via a new one-time meta gate (
trustedsigners_expanded_25211) that INSERT OR IGNOREs just the five names into the store and republishes only if something was actually added. A second boot is a byte-for-byte no-op, and an admin removal afterwards stands — the gate never resurrects a removed name. The inert 'Valve Corporation' row from 2.5.205 is untouched: it matches no real signer, fails closed, and can still be removed in Admin → Protection → Trusted signers. - Attempted and deliberately NOT added this run: Signal (its official download URL now 404s), OBS Studio, 1Password, and WhatsApp Desktop (their official URLs served "Not Found" or HTML error pages instead of signed binaries). A candidate with no vendor-served signed binary does not ship — nothing here was padded in from documentation or memory.
- Docs in the same push: admin guide gains the 2.5.211 card; the homepage is unchanged (admin-only releases never rotate it — it still leads with 2.5.210, and the fold still heads with 2.5.203).
Version 2.5.210
The Threat Guide's daily update — 3 genuinely new entries, 147 → 150, in English and Spanish alike. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged.)
- 3 new entries, all in the house format (what it is / how it gets in / how AEAntivirus stops it) with a defensible danger rating on each: Preinstalled Android malware (Critical, Spying & stealing — Bitdefender, October 8, 2026, "Midnight Mimosa": a malicious system app baked into the firmware of low-cost Android phones, mostly MediaTek models including budget brands and counterfeit flagship look-alikes, before the owner ever switches the phone on — with system-level privileges to silently install and remove apps, grant permissions, and download new code, turning the phone into an ad-fraud machine and a residential-proxy relay; a factory reset may not remove it, and a phone whose system apps keep reinstalling themselves should be returned or replaced, not merely rescanned); Verification-code forwarding ("wrong code") scam (High, Tricks, traps & nuisances — the attacker enters the victim's number on their own device, triggering a genuine six-digit registration code to the victim's phone, then a message from a hijacked friend's account claims the code "arrived by mistake" and asks for it back — forward it and the attacker registers the victim's account on their phone, locks them out, and runs the same trick on their contacts; the defense is a rule, not software: a verification code is never shared with anyone, and two-step verification (a PIN) means a code alone cannot move the account); Fake FBI / IC3 recovery scam (High, Tricks, traps & nuisances — FBI PSA I-072026-PSA, July 20, 2026, resurfacing in October reporting: after someone loses money and posts about it, an "FBI agent" or "IC3" contact reaches out on Facebook Messenger or Telegram, sometimes with an AI-generated deepfake video of a senior FBI leader, and sends a link to "update" the complaint — the link opens a look-alike IC3 site that copies the real layout, harvests name, phone, email, scam type, and loss amount, issues a fake reference number, and then claims the money has been recovered for a fee, a "tax," or wallet access; the real IC3 has no social media presence, never contacts victims on chat apps, and never asks for payment to recover funds).
- Danger totals now read Critical 12 / High 57 / Medium 67 / Low 14. Only the preinstalled phone malware rates Critical — it sits below the operating system with system privileges and survives a factory reset; the two account-and-recovery scams rate High because each ends in a taken-over account or a second loss, still requiring the victim to complete one step (forwarding the code, paying the "release" fee).
- Checked and NOT duplicated: SafePay (CISA advisory AA26-293A — a ransomware group name, not a new threat type; already rejected in 2.5.202 and 2.5.208), the iPhone Duo launch phishing sites and farm-equipment business impersonation (already inside fake-shopping-sites and marketplace-scam — a product name and a business category, not a new trick), GhostBat overlay banking malware (already inside android-overlay and banking-trojan), the Trezor real-domain phishing email (already inside real-order-email and phishing — a genuine sender channel carrying attacker text, with authentication passing), fake Meta Careers interview-booking pages (already inside job-scam, fake-meeting-invite, and aitm-phishing), and the Iron Man / YAM Store phishing-kit operation (already inside phishing and smishing — a kit supplier, not a new victim-facing type). 150 is the honest count.
- Coverage claims stay honest: entries lean only on shipped layers — Android app scanning by signature and behavior, the sensitive-permissions view, web/DNS protection, and email phishing scoring. Where the decisive counter is a habit, not software (never share a verification code, never follow a link to the complaint center — type ic3.gov yourself, return a phone that reinstalls its own system apps), the entry says so plainly and gives the advice that actually works.
- Full EN+ES parity from day one: the Spanish edition gains all 3 in the same release — same slugs, same order, same severities — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 150 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - Docs in the same push: admin guide gains the 2.5.210 card; homepage "New in" rotation (this section, 2.5.208, and 2.5.204 visible; Server 2.5.203 heads the fold and the fold summary names it). The Spanish homepage carries no "New in" archive by design (it links the changelog), so no ES rotation applies.
Version 2.5.209 (admin-only)
The memory diet: the feed pull → ingest → publish path streams in bounded chunks instead of stacking whole-corpus copies in memory. (Server-only internal performance fix — no customer-facing change; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged, and the homepage "New in" headline still tracks 2.5.208.)
- Signature publishing streams. publishSignatures used to hold the full row-object array, the full line array, the joined file text, a re-read of the written file, and the delta builder's two full line sets in memory at once. It now iterates the signature table in 2,000-row batches and writes each formatted line to the temp file as it is produced — same atomic rename, byte-identical output (pinned by a test that rebuilds the file from the table and applies the published delta to the previous file byte for byte) — hashing incrementally as it goes. The version+1 delta rides the same pass (the old file's line set is deleted as lines re-emit, so what remains is exactly the DEL set), and the delta self-check hashes the applied reconstruction instead of materializing it.
- threats.json saves stream. Every feed ingest used to re-serialize the entire threats object in memory (the production corpus is ~41 MB pretty-printed) and write it non-atomically, once per feed in a cycle. The save now writes bounded chunks to a temp file and renames it into place — the parsed content is identical, and the save is atomic for the first time.
- Provenance is collected once. The ingest used to keep the raw (value, kind) pair list, a dedupe Set, and the deduped array alive together — three copies of a big pull. One Map, deduped as indicators are noted (same first-sighting order, same rows written), replaces all three. On the same path, feed-payload validation now counts via size/length instead of spreading Sets into throwaway arrays, and the threats merge checks membership against two plain Sets instead of building ~220k prefixed string copies on every ingest.
- Measured with the production-scale harness (server/tests/scale/feedmem.js — 165k feed domains + 55k IPs + 220k meta entries, ~830k DB rows, 150k signatures, all seven big feeds mock-wired locally, RSS sampled every 100 ms through boot plus a full feed cycle: the expiry sweep, all seven pulls including CERT.PL's 125k rows, and the 155k-signature publish): peak RSS fell from 880 MB on the 2.5.208 code to 570 MB on this one under suite load, and from 758 MB to 500 MB on a quiet box — about a third off the peak either way. Post-cycle state matched on every invariant the harness checks: all seven pulls recorded identical rows in / added / duplicates, the published file holds the same 155,003 signatures at signatures v101, and provenance lands at the same 363,334 rows. Baseline after boot (~137 MB) and the settled resident set are unchanged: the diet removes transients, not the working set. The 2.5.207 investigation's NOT-proven verdict on the production restart cause stands as written — this release cuts the measured peak on the path regardless.
- Behavior is unchanged everywhere else: ingest and publish still run in small yielding transactions, the first post-boot feed sync still waits 5 minutes and pulls sequentially, the never-block guard, pull validation/quarantine, per-type expiry, and provenance recording are untouched, the client signature format and the delta chain are byte-identical, the free-plan daily snapshot still stamps from the published file, and no manifest is stamped with the server version. No new dependencies.
Version 2.5.208
The Threat Guide's daily update — 4 genuinely new entries, 143 → 147, in English and Spanish alike. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged.)
- 4 new entries, all in the house format (what it is / how it gets in / how AEAntivirus stops it) with a defensible danger rating on each: Phone-number deactivation scam (FTC impersonation) (Medium, Tricks, traps & nuisances — the FTC's October 22, 2026 consumer alert: a caller claiming to be an FTC official says a phone registered in your name and number is being used to run scams in another country, reads a fake badge number, threatens to deactivate your phone number if you do not cooperate, and transfers you to a fake police officer to "file a report" — the report harvests your ID and personal details under a secrecy instruction, and a later call claims you are under arrest for those crimes unless you send money for a "security deposit" or travel to that country); Real order-email scams (Medium, Tricks, traps & nuisances — Cisco Talos, October 21, 2026: attackers plant their own text in fields a retailer's system later prints — shipping name, billing block, order note — so a genuine order notification really sent by the retailer (the Dell and Staples campaigns Talos analyzed, ~1,400 messages) carries an expensive purchase you never made and a "helpline" number in the Ship To / Bill To block; there is no link and no attachment, the payload is the phone number, and the message passes DKIM and DMARC because the retailer really sent it); InstallFix fake AI-tool install scam (High, Tricks, traps & nuisances — Push Security via BleepingComputer, October 2026: fake install pages for AI tools, coding assistants above all — most lures impersonate Claude or ChatGPT and Codex — exploit the fact that these tools legitimately install with one terminal command; the newest pages display the vendor's real install command on screen while the Copy button puts a different command on the clipboard, and the terminal prints the genuine vendor address while decoding a hidden one; shared AI chats and artifacts that link to the install page — LLMshare — are folded into this entry as a delivery route, not duplicated as a separate one); Patched banking-app replacement (Hirudo) (High, Spying & stealing — ThreatFabric, October 14, 2026: lure apps that work as advertised show a fake "update" that installs the trojan, which talks the victim into uninstalling the real banking app and then installs a patched copy in its place — the same app with a hooking engine inside that substitutes a genuine signing certificate when the app checks, returns an empty accessibility-services list, removes screenshot protection, and relays Play Integrity checks to the attacker's server, while the trojan steals credentials and PINs, reads messages, streams the screen, and takes remote control).
- Danger totals now read Critical 11 / High 55 / Medium 67 / Low 14. Only InstallFix and the patched banking-app replacement rate High — one ends in the victim running malware with their own hands like its ClickFix/FileFix siblings, the other is a working banking trojan with remote control; the two consumer scams are real, current loss vectors that still depend on the victim completing a step (calling the number, handing over ID, sending the deposit).
- Checked and NOT duplicated: GlitchFix / ErrTraffic garbled-page lures (already inside ClickFix and the fake-update entry — a page made to look glitched, a fake browser-update or missing-font prompt, the same paste-and-run action), SafePay (CISA advisory AA26-293A, October 20, 2026 — a ransomware group name, not a new threat type; already rejected in 2.5.202), the LiteLLM release-pipeline compromise (a real supply-chain incident, already inside the supply-chain and malicious-packages entries), PHP2Shell (CVE-2026-109618 / CVE-2026-109619 — already inside the exploit entry), the calendar-invite phishing surge (already the calendar-invite-scam entry), fake traffic-ticket QR texts (already inside quishing and government impersonation), and the ScarlettGroup extortion gang (a group name for theft-only extortion, already inside data-theft-extortion). 147 is the honest count.
- Coverage claims stay honest: entries lean only on shipped layers — fileless detection, AMSI script scanning, API-sequence analysis, Clipboard Guard, email phishing scoring, web/DNS protection, Android app scanning by signature and behavior, the sensitive-permissions view, the scam-call layers, and the Senior Protection Pack's warnings. Where the decisive counter is a habit, not software (never call a number printed in an order email, never uninstall your bank app on another app's say-so, the FTC never threatens to disconnect a phone), the entry says so plainly and gives the advice that actually works.
- Full EN+ES parity from day one: the Spanish edition gains all 4 in the same release — same slugs, same order, same severities — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 147 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - Docs in the same push: admin guide gains the 2.5.208 card; homepage "New in" rotation (this section, 2.5.204, and 2.5.203 visible; Server 2.5.202 heads the fold and the fold summary names it). The Spanish homepage carries no "New in" archive by design (it links the changelog), so no ES rotation applies.
Version 2.5.207 (admin-only)
Admin-panel incident fixes, from the 2026-10-23 console reports: the panel's own poll could throw on every cycle, two of its error paths were never real, and its polling could lock the owner out. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged. Nothing customer-facing changes — the homepage "New in" continues to track 2.5.204.)
load() threw on every poll (latent since 2.5.194): the Licenses render sits behind a lazy-tab guard — if (document.getElementById('t3')) — whose body is the single following statement. The 2.5.194 plan-filter edit inserted two statements between the guard and the render, so the guard ended up guarding the filter variables and the render ran unconditionally: until the Licenses tab had been opened once, every poll cycle threw TypeError: Cannot set properties of null (setting 'innerHTML'), and everything after that point in load() (the bug-report counts/render, presets) never ran. All four lazy-tab guards of this shape (t3, t4 in load(); t5, t6 in loadThreats()) are now braced, so an inserted statement can never silently become a guard's body again; an unguarded bg.bugs dereference on the same path (threw whenever the bugs fetch failed) was fixed with them.sectionError existed only as a call: the feed-health, update-channel and webhook-deliveries cards (2.5.160) have called sectionError(...) from their catch blocks since they shipped, but the helper was never defined — any failure in those loaders threw a ReferenceError and the section silently kept its stale content. The helper now exists: a failed section says so in place and offers a Retry.- The panel could lock its owner out (the 429 storm): the Basic-auth brute-force lockout (5 failed attempts per IP → 15-minute 429) counted every unauthenticated admin-API request as a failed attempt — including requests presenting no credentials at all, which is exactly what every browser fetch looks like once an admin session lapses (the panel polls with its session cookie, not Basic headers). One poll cycle of section fetches tripped it, and from then on the panel received a plain-text
429 too many attempts body that its JSON loaders choked on. Only requests that present credentials count as attempts now. The 5-fail / 15-minute lockout for actual password guessing is unchanged, the lockout still answers before the password is checked, and the customer-login and admin-session lockouts are untouched. - An error page is not an expired session:
load()'s fetcher treated any HTML-looking body as "session expired" (login alert + page reload). A 502/429 error page therefore popped the alert and reloaded the panel during outages. The expired-session path now keys off the response status (401, or HTML on an OK response); other failures degrade to the section's empty state and the next 60-second poll retries — no alert, no reload loop, no tight retry. - Pinned by a new suite (
v25207-adminfix-test.js): the served page is executed in a DOM (jsdom) with the lazy tabs unbuilt and non-empty payloads — load() must resolve, and a forced feed-health failure must render the section's error + Retry state; over HTTP, a run of credential-less admin-API polls must return 401 and never 429, while five wrong-password Basic attempts must still trip the lockout (and a credential-less request during the lockout still gets its plain 401); the structural pins (braced guards, sectionError defined exactly once, the credentialed-attempt marker) are asserted in source.
Version 2.5.206 (admin-only)
Correction to the 2.5.205 trusted-signer expansion: one shipped name was wrong, and one real signer was missing. The default list is now 33 names. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged. Nothing customer-facing changes — the homepage "New in" continues to track 2.5.204.)
- The wrong name, corrected plainly: the 2.5.205 section below names the Steam signer as
Valve Corporation. That was wrong. After 2.5.205 shipped, every one of its 25 additions was re-verified from a fresh download with the signer resolved from the file's own signature (SignerInfo) — and the current official SteamSetup.exe contains exactly one signer certificate, reading Valve Corp.. No Valve Corporation certificate exists in it. The shipped entry matched nothing: inert, failing closed (it never excused a file), but Steam received no trust reduction while it stood. The default list now carries Valve Corp. - The missed sibling: Cisco signs Webex with two identities. The
Cisco WebEx LLC entry in 2.5.205 re-verified exact from its original artifact and stays; the Gold Webex.msi from Cisco's own binaries host signs as Cisco Systems, Inc., and that second identity is now added. (The 2.5.205 note "(not Cisco Systems)" was true of the consumer installer it was drawn from — and incomplete: both identities are real, and the list now carries both.) Defaults: 32 → 33. - Delivery to existing installs: a new one-time meta-flagged gate
INSERT OR IGNOREs just the two corrected names on first boot — including on installs where the 2.5.205 expansion already ran — and republishes trustedsigners.txt only if something was actually added. It runs exactly once: a second boot changes nothing, and an admin's later removal of either name is never resurrected. The inert Valve Corporation row is deliberately not auto-deleted — it can never match a real signer, and an admin can remove it in Protection → Trusted signers. Fresh installs seed the corrected 33 directly. Two prepared inserts — no chunked machinery, no long synchronous block (the 2.5.204 lesson). - The rest of the re-verification held: the other 24 additions re-extracted exactly (several with rotated certificates since the original pass — the names, which are what the list carries, are unchanged).
- Pinned by a new suite (
v25206-signers-fix-test.js): the defaults are exactly the corrected 33 (both new names present, Valve Corporation absent, client format contract intact); a database in the exact post-2.5.205 state (the shipped 32 including the wrong name, expansion flag already set) gains exactly the two corrected names on boot, keeps the inert row, republishes exactly once (manifest publish counter +1, file and SHA agree), and a second boot is a no-op; the 2.5.205 suite's pins are rolled to match and stay green.
Version 2.5.205 (admin-only)
The trusted-signer allowlist grows from 7 to 32 vendors. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged. Nothing customer-facing changes — the homepage "New in" continues to track 2.5.204.)
- What the list is: vendor certificate subject names (Protection → Trusted signers in the admin panel). Clients give files a heuristic-score reduction when they carry a valid Authenticode signature from a listed vendor — fewer false positives on legitimate software. The signature must still verify; the list never excuses an invalid one. Matching is unchanged: the client compares the signer's simple display name against the list (case-insensitive substring), capped at 256 entries — no semantics were loosened, and every addition is in the same name form as the original seven.
- How the additions were verified: every one of the 25 new names is the signer subject extracted from that vendor's own signed installer, downloaded from the vendor's official channel during the build and inspected byte-for-byte (PE security directory / MSI signature stream → PKCS#7 → signer certificate subject). Nothing was added from documentation, memory, or a plausible-looking guess — and the exercise proved why: Dropbox's certificate reads
Dropbox, Inc (no trailing period), Webex signs as Cisco WebEx LLC (not Cisco Systems), the Epic Games Launcher as Epic Games Inc. (no comma). The added vendors: Adobe, AnyDesk, Bitdefender, Brave, Cisco WebEx, Discord, Dropbox, Epic Games, Gen Digital (the Avast installer's signer — Avast and Norton share this parent), Git for Windows (maintainer Johannes Schindelin), Intel, JetBrains, NVIDIA, Notepad++, Opera, Oracle, PuTTY (Simon Tatham), Python Software Foundation, Spotify, TeamViewer, The Document Foundation (LibreOffice), Ubisoft, Valve, VideoLAN, Zoom. - Delivery: existing installs receive the additions exactly once on first boot (a meta-flagged
INSERT OR IGNORE expansion over the same default list the fresh-install seed uses), then publishTrustedSigners() republishes server/updates/trustedsigners.txt and bumps the manifest's trustedsigners_version through the existing channel. An admin's later removals are never resurrected. The work is 32 prepared inserts — deliberately *not* routed through the chunked ingest machinery, which exists for hundred-thousand-row feed pulls; nothing in this release adds a long synchronous block (the 2.5.204 lesson). - Deliberately not added (no verifiable binary, no entry): AMD (its CDN redirects scripted downloads to a "Download Incomplete" page), Dell / HP / Lenovo (support portals are script-gated; no clean official artifact was fetchable from the build sandbox), Realtek (distribution reachable from here carries Microsoft WHQL signatures, which evidence Microsoft's identity, not Realtek's), Samsung (no direct artifact URL exposed), WinSCP and KeePass (their download paths served HTML interstitials instead of binaries to the sandbox), McAfee (tokenized consumer download flow), and 7-Zip — whose current official installer is unsigned (verified from the bytes: no security directory), so it has no signer identity to add. Each remains addable from the admin panel the moment a real certificate is in hand.
- A number worth correcting: the
1948 seen next to trusted signers in the update manifest is the manifest's publish *counter* (it increments on every publish, including sandbox test boots — the build tree's copy read 2784), never a signer count. The live list held exactly 7 names before this release; it now holds 32. - Pinned by a new suite (
v25205-signers-test.js): a fresh boot serves exactly the 32-name set; a pre-2.5.205 database (original 7 names, no expansion flag) gains all 25 additions on boot and republishes; an admin removal followed by a reboot does not resurrect the removed name; every entry passes the client format contract (non-empty, ≤127 chars, printable, no duplicates case-insensitively); the published text file and the manifest SHA agree.
Version 2.5.204
Hotfix: the feed pipeline can no longer freeze the server. On 2026-10-22 the production site returned 502 Bad Gateway across the board after 2.5.203's feeds met production scale; this release removes the freeze class end to end. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged. No detection data, feed roster, or signature-format change.)
- What actually happened (diagnosed from the code): node:sqlite is synchronous, and 2.5.203's ingest did its work in uninterruptible blocks — the new provenance flush was a single transaction spanning every indicator in a pull (~125,000 for CERT.PL, each with an attribution read, a source upsert, and a distinct-source recount), the hash loop re-prepared its two statements for every hash, signature publishing rebuilt the whole database text in one pass, and the expiry sweep deleted its entire backlog in one transaction inside the feed tick. Worse,
maybePullMisp fired every overdue feed in the same tick without awaiting any of them, and the misp-sync job ran 90 seconds after every boot — so each restart (deploy or recycle) immediately re-entered the heaviest path with every overdue feed at once, the first-ever production CERT.PL pull included. While a block ran, every request — homepage, admin console, /api/version — queued behind it until the proxy gave up and returned 502. - Chunked, yielding ingest everywhere:
mispIngest (hash loop, cert loop, provenance flush), the admin intel feed's provenance loop, and publishSignatures now work in bounded chunks (2,000 rows per transaction) with an event-loop yield between chunks. No pull can freeze the server for more than a brief moment, and a failed provenance batch logs and lets the flush continue with the next batch instead of abandoning the pull. - Expiry sweep drains in batches: the candidate scan is capped per run (50,000 values per type) and deletes land in chunk-sized transactions with yields; a capped run deliberately does not stamp its 20-hour gate, so the next 15-minute tick keeps draining in bounded batches.
- Boot stagger + sequential pulls: the first feed sync after a restart waits 5 minutes (was 90 seconds), due feeds then pull one at a time with a 2-second gap, and a tick that finds a sync already running stands down.
- Cheaper public stats:
/api/threats-stopped — polled every 30 seconds by every homepage visitor's ticker — is now served from a 15-second in-memory cache (the counter is one monotonic row; real detection increments and the owner reset route both invalidate the cache immediately, so the cache only absorbs idle polling). /api/protection-stats keeps its existing 5-minute cache. - Homepage console error fixed: the hero animation's "pop" timeout declared a callback parameter that was never passed, so every pop threw
Cannot read properties of undefined (reading 'classList') in the browser console — the error in Austin's 2026-10-22 DevTools screenshot, present on both the English and Spanish homepages and unrelated to the 502s. The element is captured directly and guarded now, and the ticker/stats fetches check the response status before parsing, so on any failure the counters keep their last values and the next poll retries. - Pinned by a new suite (
v25204-hotfix-test.js): a synthetic 125,000-domain CERT.PL-format pull against a scratch server must complete with /api/version responsive throughout (polled latency asserted), the threats-stopped cache must serve the pre-bump value inside its TTL, the batched expiry sweep must drain seeded stale rows via the admin endpoint, and the served homepages must carry the fail-soft guards (and none of the unguarded pop timeout).
Version 2.5.203
Three new default threat feeds — and a five-part hardening of the whole feed pipeline: provenance that records every source, expiry for stale indicators, a structural never-block guard, pull validation with quarantine, and feed health that shows each feed's real contribution. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged. The client signature format does not change.)
- Three new default feeds, each live-fetched and parsed from the build sandbox before shipping (a feed that cannot be fetched and parsed here does not ship): ET Open compromised IPs (Proofpoint's free ET Open tier; https://rules.emergingthreats.net/blockrules/compromised-ips.txt — hosts known to be compromised, a different signal from the C2/attacker lists already pulled; plain text, 12h cadence; proof: 477 IPs parsed); GreenSnow scanner IPs (https://blocklist.greensnow.co/greensnow.txt — IPs caught scanning/attacking the GreenSnow sensor network; plain text, 12h; proof: 4,721 IPs parsed); CERT.PL warning-list domains (CERT Polska, the Polish national CERT; https://hole.cert.pl/domains/v2/domains.csv — the public warning list of active phishing/fraud domains, TSV with a delisting column the new
extractCertPlDomains parser honors; 24h cadence; proof: 125,336 active domains parsed, 43 delisted rows skipped). Default feeds go 13 → 16. - Evaluated and deliberately NOT added (proof in the 2.5.203 build report): abuse.ch SSLBL JA3 fingerprints (the list is abandoned — its own header says "Last updated: 2021-08-03"); Binary Defense banlist (its header forbids commercial use verbatim); Spamhaus DROP (live, but it publishes CIDR ranges the plain-IP client feed contract cannot carry, and commercial redistribution is Spamhaus's paid tier); CISA KEV (live, but CVE/vulnerability data maps to no indicator store); Tor exit nodes (not malicious indicators — a policy exclusion, not a data gap); NoThink (404, dead); CoinBlockerLists (no keyless fetch path — GitLab raw returns a sign-in page, the GitHub mirror 404s); DigitalSide (connection fails repeatedly from the build sandbox — unverifiable); the ISC suspicious-domains list (404 at both of its hosts); Phishing.Database and IPSum/FireHOL (aggregators re-listing sources already pulled — same data under a new name is not an addition, and the ~11 MB payload would ride every client's feed download).
- Improvement 1 — cross-feed dedupe with full provenance. One indicator is still stored once, but now every reporting source is recorded:
ioc_sources gains last_seen (refreshed on every sighting) and ingest bookkeeping is batched into a single transaction (flushProvenance). This fixes a latent defect found in the audit: the old per-item path returned early whenever feed_indicators already attributed an indicator to a different source, so second and later feeds were never recorded — cross-feed agreement (ioc_boost, the admin correlation view, and 2.5.0's consensus auto-quarantine, which promotes at 3+ agreeing feeds) could only ever see the first feed. First-source attribution in feed_indicators and the tp-credit rule (one credit per pull per corroborated indicator) are unchanged. The admin-configured intel feed now records provenance too (source 0), so its indicators participate in agreement and expiry like everyone else's. - Improvement 2 — expiry/decay. Feed-sourced IPs age out 30 days after the last feed listing them stops re-listing them; feed-sourced domains after 90 days. Hashes and certificate fingerprints persist (evidence, not weather). The sweep (
sweepFeedExpiry, riding the 15-minute feed tick at most once per 20h, plus an owner-only POST /api/admin/feeds/expire-sweep) removes an indicator only when every source that ever listed it is stale past its TTL — admin-entered and unprovenanced stock is always kept (when in doubt, keep). Expired indicators leave the threat store, ioc_sources, ioc_boost, and feed_indicators together; every sweep that removes something is audit-logged. Consensus promotions are deliberately not auto-revoked by expiry (an admin decision surface, unchanged). - Improvement 3 — never-block guard at ingest, structural and test-pinned. A curated protected set — AEAntivirus's own domains and the owner's business infrastructure, the platforms updates/payments/hosting depend on (GitHub, Render, Cloudflare, Stripe), the major platform domains whose blocklisting would strand customers (Microsoft/Windows Update, Google, Apple, Mozilla, Amazon, Meta), AV-vendor and feed-provider domains — matched at the root and one label below it (its service subdomains, e.g. docs.google.com), plus every hash the server itself vouches for (the known-good whitelist and every file hash in the update manifest, so our own installers can never become indicators). Enforced in
mispIngest and the intel-feed path: protected values are rejected before any merge, counted as rejections, and never recorded as provenance. This is the poisoned-blocklist lesson made structural — the google.com incident was previously fixed only at match time. Calibration from the first live CERT.PL pull: 633 of its entries are phishing pages on *individual* S3 / Cloud Storage buckets (e.g. ….s3.eu-north-1.amazonaws.com, ….storage.googleapis.com) — deep tenant hosts stay listable, because the store condemns a listed host and everything under it, never its parents or siblings; condemning amazonaws.com over one bucket would be the google.com lesson all over again, and the drill suite pins both directions. - Improvement 4 — pull validation with quarantine. Feed bodies are read through a bounded streaming reader (32 MB cap, content-length pre-check, truncation detection — a body short of its declared length fails the pull); JSON kinds parse from the bounded text (an unparseable body is a failed pull, not a scheduler exception); extracted payloads are shape-checked (an HTML error page served with a 200 is rejected; zero extracted indicators = quarantined); and a shrink guard quarantines any pull extracting under 25% of the feed's recent median (median ≥ 50, ≥ 2 prior good pulls) — the dropped-hashes failure mode (2.5.60) can no longer half-land silently. A quarantined pull ingests nothing, is recorded in feed health with its reason, and counts toward the source's consecutive failures. Fetch timeout rises 30s → 60s (the CERT.PL list is ~7 MB). Test-only
FEED_ALLOW_PRIVATE=1 relaxes the SSRF check for mock feeds, mirroring the existing TAXII_ALLOW_PRIVATE precedent. - Improvement 5 — feed health++.
feed_pull_stats gains per-pull dupes / rejected / quarantined counts (mispIngest reports skipped-as-duplicate separately from rejected-as-invalid-or-protected; the intel feed reports the same shape). The admin feed-health card now shows, per feed: latest-pull added/dup/rejected, a red QUARANTINED status, a "also reported by 1+ other feeds: N indicators" corroboration count (live from ioc_sources ⋈ ioc_boost), and a header line with the expiry sweep's last run and all-time expired total. - Two latent admin bugs fixed in the same pass:
sanitizeMispSource's kind whitelist had silently omitted the seeded kinds sslbl_ip, bambenek, cins, malc0de, and blocklistde since 2.2.0 — an admin *editing* one of those sources got a 400. The whitelist now matches the puller's dispatch exactly (plus the three new kinds), and the admin kind dropdown offers ET Open / GreenSnow / CERT.PL. Also, the URL/IP meta-attribution loop no longer does Array.includes per indicator (O(n) each) — membership is via Sets, which is what makes a 125k-domain pull practical; the old shape would have taken minutes-to-hours per CERT.PL pull. - The homepage's "15 feeds" is now a live count. The trust strip hardcoded 15, a figure that counted nothing (the seeded defaults were 13).
liveFeedCount() now counts enabled feed sources (not precision auto-paused) + the admin intel feed when configured + the MalwareBazaar staged import when its key is set — the same no-drift pattern as the Threat Guide count — and renders into both homepages (EN + ES) and the Everything-included feature line. The Help page's feed sentence names the new feeds and drops the stale numeral (its renderer takes no count context). - Publish path unchanged where it matters: pulls still flow to signatures + manifest through the same atomic temp+rename publish; the free-plan daily signature snapshot cadence is untouched; no client file, manifest binary, or signature format changes. Ops note: the first CERT.PL pull adds ~125k domains to the threat store in one ingest — that is the feed working as designed, and the expiry policy above is what keeps it from only ever growing.
Version 2.5.202
The Threat Guide's daily update — 4 genuinely new entries, 139 → 143, in English and Spanish alike. (Server-only release; Windows client 2.5.151, Android 1.39.0, and Linux 1.29.0 are unchanged.)
- 4 new entries, all in the house format (what it is / how it gets in / how AEAntivirus stops it) with a defensible danger rating on each, all in the Tricks, traps & nuisances group: CrashFix browser-crash scam (High — ClickFix's newest sibling: a fake ad-blocker extension overloads the browser until it crashes, about an hour after install, then shows a "security warning" whose offered fix is the same Run-box paste as ClickFix, delivering a remote-access tool or stealer; Huntress / Trend Micro reporting, still circulating in October 2026 coverage); Election scams (Medium — the FBI's October 15, 2026 pre-midterm PSA I-101526-PSA-3: "campaign investment pools" promising a payout on donations, fake PACs pocketing real donations, campaign merchandise paid for and never shipped, and "you are not registered" texts linking to look-alike registration pages that harvest personal information, with AI-made candidate voices and images raising the polish); Brushing scam (mystery package) (Medium — a parcel you never ordered, no sender named, whose enclosed QR card ("find out who sent it") leads to card/personal-data harvesting or malware, on top of the fake "verified purchase" review fraud the shipment itself exists for; FBI IC3 + U.S. Postal Inspection Service warnings); Mobile-wallet enrollment scam (Medium — ordinary phishing harvests card details, then the "purchase confirmation" code the bank texts is actually a wallet-enrollment code: sharing it adds the victim's card to the scammer's Apple Pay / Google Wallet, whose phone can then tap-to-pay as the victim; Krebs and Bitdefender provisioning-fraud reporting).
- Danger totals now read Critical 11 / High 53 / Medium 65 / Low 14. Only CrashFix rates High — it ends in hands-on malware installation like its ClickFix/FileFix siblings; the other three are real, common loss vectors that still depend on the victim completing a step (scanning, donating, sharing a code).
- Checked and NOT duplicated: VHDX virtual-disk delivery (already inside the disk-images entry, which names .vhd virtual drives, and the LNK entry), jury-duty arrest threats (already inside government impersonation, verbatim), parking-meter QR stickers (already inside quishing, verbatim), sextortion-relief / "identify your blackmailer" recovery offers (already the recovery-scam entry's second-bite pattern — the FTC's October 13, 2026 Digital Forensics case), and SafePay (CISA advisory AA26-293A, October 20, 2026 — a ransomware group name, not a new threat type; the guide's ransomware / double-extortion / RaaS entries already cover its behavior). 143 is the honest count.
- Coverage claims stay honest: entries lean only on shipped layers — fileless detection, AMSI script scanning, API-sequence analysis, Clipboard Guard, the browser extension auditor, email/message phishing scoring, web/DNS protection and the threat feeds, and the SMS Phishing Filter add-on. Where no software counter exists (a cardboard box, a donation decision, a bank code read aloud), the entry says so plainly and gives the advice that actually works — never scan a mystery package's QR code, a donation is never an investment (check the FEC's own records; register only via vote.gov typed in yourself), and read the whole bank message before typing any code, because the code names what it approves.
- Full EN+ES parity from day one: the Spanish edition gains all 4 in the same release — same slugs, same order, same severities — so no group falls back to English. The homepage strip, About tile, emergency one-pager, and the
/api/threat-guide payload keep rendering the count live from the data (they read 143 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged (still the current month). - One-line rider fixed: the server's
PLATFORMS table still named Linux 1.28.0 after Linux 1.29.0 shipped (the pages read the manifest, so customers already saw 1.29.0; the table now agrees at 1.29.0). - Docs in the same push: admin guide gains the 2.5.202 card; homepage "New in" rotation (this section, 2.5.201, and 2.5.200 visible; Server 2.5.199 heads the fold and the fold summary names it). The Spanish homepage carries no "New in" archive by design (it links the changelog), so no ES rotation applies.
Version 2.5.151
The Threat Guide moves into the app, game mode holds scheduled scans, repeat scans get faster, text gets a size setting, and updates get smaller. (Client-only release; the server stays 2.5.201, Android 1.36.0, and Linux 1.28.0 unchanged.)
- Threat Guide in the app: the full guide the website shows (139 entries, plain English, English or Spanish) now opens inside the Windows app — the About page and the Threat Types page both carry a Threat Guide button. The app downloads the guide data from the server, refreshes it at most once a day, and keeps its saved copy for offline reading. Detection popups explain themselves: a What is this alert? section written for the exact detection name, and — where the server data links that detection to one guide entry — a button that opens the guide at that entry. The links ship in the server's data; the app hardcodes none of them.
- Game mode holds scheduled scans: while silent/game mode is on (switched on by hand or triggered by a fullscreen app), a due scheduled scan now waits instead of interrupting — the same treatment the battery gate already had, 24-hour cap included. A held scan is never dropped: it runs as soon as the mode turns off. Manual scans are never held. The Status page shows the mode's state — Off, On (manual), or On (fullscreen app). If the app window isn't running, no state is published and scheduled scans run on time — a scan can never get stuck waiting on a mode nobody can see.
- Faster repeat scans: a file the full scan pipeline already judged clean is remembered by its fingerprint (SHA-256), stamped with the signature version that judged it. On later scans the fingerprint is checked only after every exclusion and detection stage that outranks it, so a remembered file can never hide a detection — and the moment signatures update, every remembered verdict expires at once. The scan page and Status page show the count as skipped as unchanged, zeroed at every scan start like the other counters. Quick and Full scans share the path and both benefit.
- Text size setting: Settings gains Text size — 100%, 125%, or 150% — for the main window's pages and buttons, applied when the app restarts. Honest scope: a few deep-utility pages with dense fixed layouts (Boot Guard, Ransomware Rollback, Heuristic Tuning, Cloud Defense, Threat Intel, Advanced Heuristics, Threat Types, Parental Controls, Password Manager, System Optimizer), the Threat Guide window, and the setup wizard keep their current size rather than risk clipped text.
- Smaller updates (delta patches): starting with this release the updater keeps a copy of the installer it ran. When a later release publishes a patch, an up-to-date install downloads just the patch, rebuilds the new installer from its saved copy, and verifies the result's fingerprint against the manifest before anything runs — any failure quietly uses the normal full download instead. One honest limit: 2.5.150 installs have no saved copy yet, so this release itself arrives as a full download and saves its copy for next time.
Batch 12 release A (server-only; the Windows client stays 2.5.122):
- Forgot password — both halves (there was NO self-service reset anywhere before this release). New
password_resets table stores only a SHA-256 hash of each token (the raw token exists solely in the link), with expiry, single-use marking, source and requesting IP. POST /api/forgot-password answers with one byte-identical generic response for every input — known email, unknown email, malformed input, rate-limited — so it can never enumerate accounts; for a real account it invalidates earlier unused tokens, mints a 1-hour token and emails it through sendMail with the new password_reset template (editable in the admin template editor). Caps: 3 tokens/hour per account and per IP, counted from the token rows. POST /api/reset-password redeems with the registration password policy (8+ chars + breached-password check, shared with 2.5.147's change flow), marks the token used, revokes ALL of the account's sessions, sends the "your password was reset" security email and writes an audit entry. SMTP honesty: with no SMTP configured nothing can be delivered, but the request is still recorded and the response stays generic — and the /forgot page copy says delivery depends on mail being set up instead of promising an email (support fallback named on the page). Admin half: POST /api/admin/accounts/:id/reset-link (owner-only via the ADMIN_OWNER_ONLY denylist, next to refunds/merge/global-reset) mints a 24-hour token through the same machinery and returns the full URL exactly once; the customer Billing view gains Generate reset link with a copy-friendly read-only field; generation invalidates prior tokens and is audit-logged. Suite: v25151-batch12r1-test.js proves enumeration-identical bodies, single use, expiry, session revocation + old-password rejection, admin link invalidating the email token, the 4th-request cap, and the SMTP-absent path. - Pause subscription (the license goes inactive for the pause — Austin 2026-10-16: "Don't leave protection on during a pause"). A Stripe-billed license on My Account gains Pause billing (1, 2 or 3 months) → Stripe
pause_collection {behavior:'void', resumes_at}: charges are *skipped*, not deferred, and Stripe resumes billing on its own at the end. The pause takes effect at the end of the current paid period, read from the Stripe subscription itself: the paid remainder stays fully active, and the pause window runs [pause_from, paused_until) — mirrored on the license (licenses.pause_from / licenses.paused_until, shown as "Pause scheduled — starts <date>" before the window and "Paused until <date>" inside it, on the account card and in the admin Licenses list) and re-synced by the customer.subscription.updated webhook (Stripe is the source of truth; our own endpoint also writes the mirror immediately). Inside the window the license is INACTIVE: heartbeats empty the add-ons list (the existing expired-license rule) and licenseStateForClient reports the distinct status paused — clients turn protection off for the pause and label it "Subscription paused", never "expired" (no grace clock is emitted for a pause). Add-on subscriptions are not paused at Stripe; their entitlement simply stops flowing during the window and returns on the next heartbeat after it ends. Resume now clears the Stripe pause and the mirror immediately (it also cancels a merely scheduled pause). Rails: refused for past_due / canceled / already-paused-or-scheduled licenses, for licenses with no Stripe subscription, and for a second pause within a rolling 12 months (history in the new subscription_pauses table) — each refusal is a clear message, nothing partial. Cancel during a pause keeps working through the existing flows. ToS §subscriptions describes the pause terms, including that protection is off during the pause. Suite: v25151-batch12r1-test.js runs the full cycle against the Stripe test mock: paid period still active, effective window → add-ons emptied + status paused, resume restores both, webhook-driven mirror sync both ways. - Email preference center. The Security tab's consent toggles are gathered into ONE Email preferences card with a single save: threat alerts (
accounts.threat_alerts), family digest opt-out (accounts.family_digest_optout), monthly fleet report opt-out (accounts.fleet_report_optout), newsletter (newsletter_subs by the account's email) and outage alerts (status_alert_subs by email). GET/POST /api/account/alert-prefs is extended — no new source of truth; the POST writes each underlying store and reports applicable so the card only shows toggles that mean something for the account (family digest only for family owners, fleet report only for report-eligible tiers). The legacy per-toggle fields keep working for the existing UI call sites. Suite asserts each toggle flips its actual table row in the DB (not just the response), that inapplicable toggles are absent from the served page, and that a save leaves the account's other preferences untouched.
Version 2.5.201
The Threat Guide becomes data the app can read, the update channel learns delta patches, and Android gains two permission-audit checks. (Server-only release; Windows client 2.5.150, Android 1.36.0, and Linux 1.28.0 are unchanged.)
- Threat Guide API:
GET /api/threat-guide is public (no sign-in — the content is already the public Threat Guide page) and returns {version, count, groups, alerts} as JSON, built per request from the same guide data the website renders — one source of truth, nothing copied. ?lang=es serves the Spanish edition (groups and alert texts alike). The alerts list structures the 14 "What our alerts mean" paragraphs (15 detection names — ML.Behavior.Malicious and ML.Behavior.Suspicious share one paragraph and carry its text together), each with a slug ONLY where one guide entry unambiguously matches the detection (ML.Behavior.Ransomware → ransomware, Keylogger.Detected → keylogger, Rootkit.Hidden → rootkit, Fileless.Threat → fileless, Behavior.LOLBin → lolbins, Behavior.OfficeMacro → macro-malware, Clipboard.Hijack → clipboard-hijacker, Defense.HostsFile → dns-hijacker, Ransomware.Behavior.Canary → ransomware); the rest are null rather than a guess (Ransomware.Behavior.Sabotage included — shadow-copy destruction is covered by both the ransomware and wiper entries). Responses carry Cache-Control: public, max-age=3600. This powers the in-client Threat Guide in a later client release. - Delta update publishing (serving layer): the
/updates/proxy/:file allowlist now also serves AEAntivirus-Setup.patch and AEAntivirus-Setup-x64.patch (same name-generic repo mapping: server/updates/<file>), and the /updates/manifest.json route rewrites delta.patch_url / delta_x64.patch_url to the proxy in the RESPONSE only when they start with /updates/ — the disk manifest is never modified, unknown or absent delta fields pass through untouched, and the manifest's client version stays 2.5.150 (server/client independence). Inert until a client build publishes patch files. - Two new Android Tech tools checks:
diag-accessibility (Accessibility services — every accessibility service enabled on the phone, name + package) and diag-overlay (Draw-over-apps permissions — every app holding the draw-over-other-apps grant). They join the global diag tool list so every validation path knows the names (command validation, diag-cancel targets, schedules, progress/result ingest), the friendly-label tables EN + ES, the admin panel (order, labels, summaries, detail renderers, explainers) and the printable-report summary — but the per-platform matrix offers them ONLY to Android installs (now 10 checks). Windows keeps its 49-check catalog exactly (its matrix set is the global list minus the Android-only pair), and Windows/Linux/macOS queue attempts are refused with the platform reason, as are cross-platform result posts. The Android Full checkup keeps its 8-tool intake set; the audits queue individually. The /tech-services pages (EN + ES) now say 10 Android checks and name the pair. Android client collectors ship in a later client release. - Docs in the same push: admin guide gains the 2.5.201 card; homepage "New in" rotation (this section, 2.5.200, and 2.5.199 visible; Server 2.5.198 heads the fold and the fold summary names it). The Spanish homepage carries no "New in" archive by design (it links the changelog), so no ES rotation applies.
Version 2.5.200
The Threat Guide grows again — 19 genuinely new entries, 120 → 139, in English and Spanish alike. (Server-only release; Windows client 2.5.150, Android 1.36.0, and Linux 1.28.0 are unchanged.)
- 19 new entries, all in the house format (what it is / how it gets in / how AEAntivirus stops it) with a defensible danger rating on each. The yield is the honest remainder of a slug-by-slug audit: thirty candidate topics were checked against all 120 existing entries, and twelve of them — formjacking (already the web-skimmer entry), OAuth consent phishing, pass-the-cookie (already session hijacking), browser-in-the-browser, juice jacking, watering-hole attacks, pig butchering, task scams (already the job-scam entry's coverage), gift-card *payment* scams, browser notification scams, domain typosquatting, and fake-CAPTCHA tricks beyond ClickFix/FileFix — were already in the guide by name or inside an existing entry, so they were NOT duplicated. What shipped: Spying & stealing — overlay attacks (Android), accessibility-service abuse, QR login jacking, hidden cameras, IMSI catchers (fake cell towers). Money — gift-card draining (distinct from the gift-card payment scam: the card is emptied before you buy it), crypto ATM scams, gold-bar courier scams, check washing & mail theft, home title (deed) fraud. Tricks — utility shutoff scams, fake debt-collector scams, money-mule recruitment, look-alike (homograph) domain phishing, tabnabbing. Stealth & network — DNS rebinding, relay attacks (NFC, RFID & keyless entry), malicious PDF documents, SVG phishing lures.
- Danger totals now read Critical 11 / High 52 / Medium 62 / Low 14. The new highs are the two Android session thefts and QR login jacking; the rest sit at Medium (and tabnabbing at Low), because each is a real, common loss vector that still depends on the victim completing a step.
- Coverage claims stay honest: entries lean only on shipped layers — signature + behavioral scanning, the sensitive-permissions view and install-time scanning on Android, USB Defense, web/DNS protection and the threat feeds, the scam and link checkers, the Scam-call Blocker and the Senior Protection Pack. Entries with no software counter (IMSI catchers, hidden cameras, gift-card draining, check washing, deed fraud, relay attacks) say so plainly and give the advice that actually works — Faraday pouches, county property-fraud alerts, gel pens, Informed Delivery.
- Full EN+ES parity from day one: the Spanish edition gains all 19 in the same release — same slugs, same order, verified group-by-group, so no group falls back to English and every count matches. The homepage strip and About tile keep rendering the count live from the data (they read 139 with no page edits), and the guide page's "Last reviewed: October 2026" stamp is unchanged.
- Docs in the same push: admin guide gains the 2.5.200 card; homepage "New in" rotation (this section, 2.5.199, and 2.5.198 visible; Server 2.5.197 heads the fold and the fold summary names it).
Version 2.5.199
The Threat Guide grows to 120 entries, every entry gains a danger rating, and every count on the site renders live from the guide data. (Server-only release; Windows client 2.5.150, Android 1.36.0, and Linux 1.28.0 are unchanged.)
- 36 new entries (84 → 120), all in the house format (what it is / how it gets in / how AEAntivirus stops it), tied only to real shipped engines — and where AE has no counter (physical card skimmers, Bluetooth trackers), the entry says so plainly and gives the protective advice instead. Classic: BadUSB & keystroke-injection devices, physical-access ("evil maid") attacks. Money: crypto wallet drainers, ransomware-as-a-service, theft-only extortion, crypto address poisoning, card skimmers & shimmers. Spying: Bluetooth tracker stalking, webcam & microphone spying, device-code phishing. Tricks: adversary-in-the-middle (AiTM) phishing, AI chatbot impersonation, callback phishing, payment-app scams, romance scams, family-emergency ("grandparent") scams, Medicare & health-benefits scams, fake online stores, fake shared-file lures, recovery scams, prize & lottery scams, social-media giveaway scams, fake & malicious phone apps, fake PC cleaners & driver updaters, calendar-invite phishing, refund scams, fake meeting invites, the FileFix paste trick, email bombing. Stealth: BYOVD, HTML smuggling, malicious software packages, zero-click exploits, ARP spoofing, abused remote-access tools, proxyware. The Spanish edition (
definitions-es.js) gains all 36 in the same release — same slugs, same order, verified group-by-group, so no group falls back to English. - Danger rating on all 120 entries: each entry tuple gains a sixth element —
low / medium / high / critical — rendered as a small colored tag beside the entry name (grey Low, amber Medium, orange High, red Critical), with translated labels in Spanish (Bajo / Medio / Alto / Crítico). Firmware/bootkit/ransomware-class threats sit at Critical (11 entries); nuisance-class entries at Low (13). Because the tag text is part of the card, the guide's existing live search now also filters by rating words — no schema change beyond the tuple element, and no new filter UI was needed (the search box has existed since 2.5.113). - Freshness line: the guide page now shows "Last reviewed: October 2026" / "Última revisión: octubre de 2026" under the intro, in both languages.
- No-drift counts: the homepage trust strip no longer hardcodes the guide size — the home routes pass
definitionsView.guideCount (computed from GUIDE_GROUPS, the single source of truth) into both home views, and the emergency one-pager (/emergency — "I'm infected") takes the same live count. The remaining "84" occurrences are historical release prose (homepage "New in" archives, older README sections, older admin-guide cards) describing the guide as it was at those releases, and stay as history. The Scam of the Week derivation from the same module-scope data is untouched and verified. - Docs in the same push: admin guide gains the 2.5.199 card; homepage "New in" rotation (this section, 2.5.198, and 2.5.197 visible; Server 2.5.196 heads the fold and the fold summary names it).
Version 2.5.198
My Account answers its two questions on sight: what state is each of my plans in, and is each device actually protected? (Server-only release; Windows client 2.5.150, Android 1.36.0, and Linux 1.28.0 are unchanged.)
- At-a-glance status strip: the top of My Account (above the tabs) now carries one row per license the account owns — platform + plan, its live state, and the state's action button inline. States are computed with the *same* effective-entitlement helpers the clients' heartbeats use (one implementation — the page can never disagree with the app): a current paid plan reads Renews <date> with a Manage button (the existing Stripe billing-portal path); a Free license inside its one full trial reads Free trial — N day(s) left (server end, ceiling, minimum 1) with Upgrade; a Free license that hasn't used its trial offers Start 7-day free trial (the 2.5.195 session-auth endpoint) + Upgrade, and one that has reads plain Free plan with Upgrade; an expired license in the Free fallback reads On Free since <date> — your plan ended with Renew pre-scoped to that license's plan on the subscribe page; a paused subscription reads Paused until <date> — distinct from fallback, never shown as on-Free. Fallback rows sort first, then a running trial, then the rest. A license with no devices still gets a row; an account with no licenses sees the existing Get Free card in the strip's place (rendered once, there instead of in the Licenses tab).
- Per-device protection state: each device row on a license card gains a plain line built only from that install's own stored heartbeat data — the same fields the admin install view and the family digest read (last check-in, client version, the four module states + when state was last reported): Protected · Last checked in <relative time> · v<version> when the last report has every protection on; Needs attention · … when anything is off; and when the last check-in is older than 48 hours the line claims no state at all — Hasn't checked in since <date>. A device that never reports module state (Android/Linux builds) shows its check-in with no verdict rather than a painted-on green. Fixed customer-safe phrases only, English + Spanish, relative times computed server-side; removed devices stay absent, as before.
- Docs in the same push: admin guide gains the 2.5.198 card; homepage "New in" rotation (this section, 2.5.197, and 2.5.196 visible; Server 2.5.195 heads the fold and the fold summary names it).
Version 2.5.197
The website gets a fresh coat of paint. (Server-only release; Windows client 2.5.150, Android 1.36.0, and Linux 1.28.0 are unchanged.)
- Hero product panel: the homepage hero now carries the AE wordmark over a soft radial cyan glow (a CSS gradient echoing the logo's own glow — no new image assets), the existing headline, copy, and CTAs beside a product panel rendered in HTML/CSS in the Windows client's own design language (palette rgb(7,11,18) / rgb(13,19,29) / rgb(232,240,250) / rgb(0,229,255)): a Status card using the client's real strings — "You're protected", "All protection running normally", the Core Protection rows (Real-time protection, Firewall, Web Protection, Advanced Real-Time Protection), "Run Smart Scan", "Lightweight - low CPU and memory usage" — plus a slimmer phone-shaped panel echoing the Android app's Status tab ("PROTECTION ON", "Scan now", its tab names). It is page presentation in the product's own styling, deliberately not labeled or captioned as a screenshot; no string in it is a claim the apps do not make, and the post-2.5.189 wording holds throughout.
- Everything included, tamed: card copy is unchanged; a render-time polish in the home views (the shared
featPolish helper in server/views/util.js) wraps each catalog card title's leading emoji in a consistent icon chip and turns the bare <span class=tag> marker into a proper pill (tag pill) — all 215 EN + 197 ES catalog cards verified chipped by the release suite. Cards render in a true two-column grid inside each group (one column under 760px). The "Jump to:" line (EN) and a new "Ir a:" line (ES) are now a sticky sub-navigation bar that rides under the sticky header, with scroll-margin-top on the group anchors so targets never land hidden underneath; the groups were already <details> and stay collapsible. - Real-numbers trust strip: a slim band directly under the hero in both languages carrying exactly four verifiable figures: 15 threat-intel feeds · the Threat Guide entry count, rendered live from the guide data itself since 2.5.199 (84 entries when this strip shipped) · 49 technician checks · "Threat signatures vNNN", where NNN is rendered server-side from the same
updates/manifest.json the update channel serves (the release suite pins the strip's number to the served manifest's sigs_version; if the manifest cannot be read, the figure is omitted rather than invented). No other numbers were added. - Sticky header CTAs: the shared site header — already sticky under the ticker — gains Get Free (/subscribe) and Download (/download) buttons in the standard button style on every public page, in both the signed-out and signed-in nav; the wordmark-only brand from 2.5.189 is unchanged, and the admin console keeps its own nav.
- Docs in the same push: admin guide gains the 2.5.197 card; homepage "New in" rotation (this section, 2.5.196, and 2.5.195 visible; Client 2.5.150 heads the fold and the fold summary names it). Homepage weight: the EN page grows by only the new markup/CSS (measured in the release notes of the build; the shared CSS is the bulk of it and is amortized across every page).
Version 2.5.196
The Rescue USB boots UEFI machines — Secure Boot included — and refreshes its signatures at boot when a wire is plugged in. (Server + rebuilt rescue image only; Client 2.5.150 and Android 1.36.0 are unchanged, and the Linux scanner stays 1.28.0.)
- UEFI boot path (Workstream A): the rescue ISO now carries BOTH El Torito boot entries in one catalog — BIOS/isolinux (platform 0x00) plus UEFI (platform 0xEF) pointing at an EFI system partition image (
boot/grub/efi.img, FAT) holding EFI/BOOT/BOOTX64.EFI (Ubuntu's Microsoft-signed shim), grubx64.efi (Canonical-signed GRUB) and mmx64.efi, with the same grub.cfg (search-and-chain to the ISO's /boot/grub/grub.cfg, self-sufficient fallback menuentry) in the ESP and on the ISO. The kernel it loads is the Canonical-signed Ubuntu generic kernel, so the full chain verifies under Secure Boot — all three signatures checked from the shipped bytes with sbverify. The build (linux/rescue/build-rescue-iso.sh) prefers this signed chain and only falls back to an unsigned grub-mkimage GRUB when the signed packages are unavailable, printing SECUREBOOT=signed|unsigned either way; this build is signed. One config subtlety kept from the design: the ESP never contains /boot/vmlinuz or /boot/grub/grub.cfg, so GRUB's search --file unambiguously pins $root to the ISO. Verified from the ISO bytes: catalog lists both entries with the right platform ids and image paths, the extracted ESP parses as FAT and every file in it is byte-identical to its signed source, and the initrd payload is byte-for-byte the verified staging set. It still cannot be boot-tested in the build sandbox (no /dev/kvm) — the first real-PC boot checks, now spanning BOTH firmware types, remain the shop's step; the /rescue copy no longer claims a BIOS/CSM limitation and states the Secure Boot support that is actually in the image. - Signature refresh at rescue boot (Workstream B):
rescue-scan.sh now runs rescue-sigs-refresh.sh under a 20-second outer timeout before scanning. Wired interfaces only: the first NIC with link comes up, DHCP gets bounded retries (in-image udhcpc.script applies the lease and writes /etc/resolv.conf), then the in-image rescue-fetch (a small libcurl tool compiled with the image; TLS verification ON against the CA bundle staged in the initrd; connect 5 s / total 6–10 s bounds) pulls /updates/manifest.json and /updates/signatures.db from the existing public update channel — no new server endpoints. The fetched DB is staged in RAM and swapped in only when its SHA-256 matches the manifest's sigs_sha256 (and the identical-set case is detected by hash, so a current snapshot is never needlessly replaced). The built-in snapshot's channel version is stamped at build time (signatures.version, read from the manifest beside the signatures source). ANY failure — no link, no lease, network/TLS error, hash mismatch, timeout — falls back to the snapshot with no drama. Exactly one line is printed on the boot screen and appended to AEAntivirus-Rescue-Report.txt: Signatures: updated at boot (vN) or Signatures: built-in snapshot (vN). No Wi-Fi configuration exists in the image; the scan itself never needs the network. The refresh needed wired-NIC drivers, so the initrd module set gained a deliberate consumer-wired subset (Intel e1000/e1000e/igb/igc, Realtek r8169, Broadcom tg3/bnx2, virtio-net, USB NICs, phy/mdio) — the wireless tree and datacenter NIC families stay out. The build also prunes the fibre-channel / iSCSI / legacy-HBA module families, packs the initrd with xz and keeps the EFI partition at 5 MiB, holding the image at 37,748,736 bytes (SHA-256 1aaa3225953e0164622970238aae897aebfb7dc356050701d46f62836f6819aa). Tested by linux/rescue/test-rescue-image.sh (22 checks: structure + chroot runs of the shipped payload — good-payload swap, corrupt-payload rejection, unreachable fallback, EICAR quarantine smoke), re-run green on the final trimmed image; the identical refresh tooling was also run end-to-end against production from inside the image during the release build (the live database was fetched, hash-verified against the live manifest, and swapped in). - Copy in the same push: /rescue EN + ES (dual boot + Secure Boot, wired refresh, new size + SHA-256), the FAQ rescue answers EN + ES, the homepage "New in" rotation (this section visible; Android 1.36.0 now heads the fold and the summary names it), and the admin guide (new 2.5.196 card; the 2.5.195 card's rescue paragraph brought current). GUIDE_TEXT is untouched — no client build.
Version 2.5.195
Try everything for 7 days from Free, expired licenses step down to Free instead of switching off, and the admin dashboard can finally see the Free funnel. (With Client 2.5.150 and Android 1.36.0.)
- One-tap 7-day full trial from the Free plan: a Free license may start ONE 7-day full trial, ever — stamped on the license row (
free_trial_used_at / free_trial_ends_at) at start, refused with "trial already used" (409) on any second start. Two entry points run the same server logic (startFreeTrial) and write the same audit entry (license.free_trial_start): the app banner (device-auth POST /api/free-trial/start) and My Account (session-auth POST /api/account/free-trial/start). While the window is open — computed from the stamps at heartbeat/entitlement time, no cron — the device's EFFECTIVE entitlements are the full paid base plan: protection modules unforced, scheduled scans allowed, LIVE signature versions, and technician services unlocked at BOTH gates (admin queue validation and command dispatch now check the effective subset, not the raw plan). Paid add-ons are NOT included — they remain separate purchases; checkout works normally during a trial (no free-refusal), and an add-on bought during a trial stays owned after, on its own terms. On expiry everything recomputes to the Free subset (daily snapshot sigs, tech refused again, trial button gone permanently). Upgrading during the trial converts the same row in place through the existing webhook path; the trial stamps stay on the row as the funnel's durable record (inert on a paid license) and the upgrade audit entry notes the supersession. - Expired licenses land on the Free subset (Austin's rule, 2026-10-21): modeled as an entitlement state, not a plan rewrite — the license row keeps its real plan and status, and effective entitlements compute to the Free subset (scans, quarantine, basic real-time protection, daily snapshot sig cadence) whenever a paid OR trial license is past plan-end + the existing 72-hour grace, until renewal. The grace itself is unchanged (full paid protection through it; only the landing changes, from OFF to Free). Renewal through the existing webhook path restores full entitlements on the next heartbeat. Add-ons are separate purchases with their own paid status: base-plan flags and add-on entitlements are computed separately, so an independently-active add-on STAYS entitled during fallback (the 2.5.194 heartbeat forced
addons: [] only for plan=free, which owns none). The heartbeat/register payloads carry fallback_free: 1 for this state (alongside trial_active / trial_ends_at / trial_eligible for the trial). Subscription PAUSE is untouched and distinct: paused still means protection OFF during the pause, and a paused license is never fallback — the suite pins both directions. - Free plan funnel (admin):
GET /api/admin/free-funnel + a dashboard card computed from the license/install tables (the audit file is capped, so the durable stamps carry the counts): Free licenses issued total + last 30 days (from the new was_free stamp, backfilled for existing free rows, so an in-place upgrade never erases an issuance), active Free installs (heartbeat within a day — the same "active" definition as the Accounts view), full-trial starts, Free → paid upgrades, conversion shown with both numbers ("1 of 4 (25%)"), and licenses currently in expired-fallback as their own line. The release suite seeds the full arithmetic — 4 issued, 3 trial starts, 1 upgrade, 3 active installs — and asserts the endpoint's numbers exactly, including the fallback count at two different moments (2 while both fixtures sit in fallback, 1 after one renews out of it). - Docs in the same push: Terms gain the one-trial paragraph in §2b and a "When a paid plan ends" step-down paragraph in §3 (the §2 trial sentence no longer says protection stops — it steps down to Free after the grace); FAQ gains the free-trial question and the expiry answer now describes the step-down (EN + ES, JSON-LD from the same list); /subscribe's Free card names the 7-day trial; My Account's Free card carries the trial button / running note / ended note; the admin guide documents the funnel, the trial, and the fallback end to end.
Version 2.5.150
The Windows app runs the trial and the fallback.
- Trial states on the Free banner: while eligible, the Free banner's button reads "Try everything free for 7 days" and starts the trial in one tap (device-auth POST to the server; the heartbeat carries the trial state so the app never guesses). During the window the banner reads "Full trial — N day(s) left" (days computed from the server-provided end, ceiling, minimum 1). On expiry the banner returns to plain Free with a one-line "Your trial ended — you're still protected on Free" note. No expired banner, no payment popup, and protection never drops below the Free subset at any transition — the client's effective-free check is now
(free && no live trial) || expired-fallback, shared by the module gates (anti-ransomware, email protection, scheduled scans) and the server-policy enforcement. - Expired-fallback banner: a license past plan-end + grace shows "Your plan ended — you're on the Free plan until you renew" with a Renew button. Protection stays ON at the Free subset (the protection gate treats fallback as protected), and the once-daily payment popup stops in this state — the banner and the renewal emails carry it. During the 72-hour grace the existing countdown UX is unchanged; a paused subscription still shows paused, never fallback.
- Reset the global threats counter (Austin 2026-10-16: "how do I clear the global threats" — he couldn't; now he can). The all-time Threats stopped total on the public homepage is a standalone monotonic counter (the
counters row behind GET /api/protection-stats), incremented once per recorded detection — it is NOT derived from detections or installs, so the per-device resets (2.5.144/2.5.146) never moved it and no reset existed anywhere. New owner-only POST /api/admin/stats/reset-threats (in the ADMIN_OWNER_ONLY denylist alongside refunds/backup/merge, behind requireOwner) requires the typed confirmation word RESET in the body, zeroes the counter (insert-or-update, so it works even if the row is absent), invalidates the 5-minute public-stats cache so the homepage band shows 0 immediately, and audit-logs the reset (stats.global_threats_reset) with the previous value; the response reports {ok, previous, now: 0}. Scope is the counter only — no detection rows are deleted and no per-device counter changes; counting resumes from the next recorded detection. UI: a Global threats counter card in System → Data cleanup showing the current total with a Reset to zero button (explain → type RESET → done). Suite: v25150-global-threats-test.js. - Server-only release; the Windows client stays 2.5.122 and Linux/Android are untouched.
Version 2.5.194
The Free plan — $0 forever, one device, core protection — plus the custom-rules cap fix (with Client 2.5.149 and Android 1.35.0).
- A new license plan,
free: $0 forever, no Stripe object, no expiry, no grace machinery. Customers issue it themselves from My Account → Get Free, choosing the platform (Windows, Linux, or Android) at issuance; the key shows on My Account labeled Free, like every plan. One Free license per account, one device per license — a second issuance is refused (409) and a second device on a taken seat is refused with a plain message. Removing the device tombstones it as usual and frees the seat for a *different* device (the replaced-PC case). - Free is its own plan on all three platforms. The 2.5.158 rule (Android installs only ever get the Mobile plan) still governs paid/trial assignment; a free-plan Android registration is accepted, and the boot mismatch audit skips free licenses.
- What Free includes: quick + full scans, real-time file protection, quarantine & removal, and signature updates on a daily snapshot. What it excludes — enforced, not just hidden: the anti-ransomware module + rollback vault, web protection, email protection, firewall extras, all paid add-ons (checkout is refused with an upgrade message), all technician services, scheduled scans, and multi-device.
- Signature cadence: a server-side daily pointer records the signature version as of the most recent 00:00 ET rollover (stamped at boot, hourly, and lazily at each rollover). Free installs that identify themselves on the update check are answered with the snapshot —
sigs_url points at the daily DB copy — while paid installs keep the live version. The pointer is a separate stored value and is never written into the manifests, so the 2.5.187 atomic publish coherence is untouched. A free device still updates every day; it is never frozen at its install-day version. - Technician services refuse Free at two layers: admin queue validation (Tech tools checks, fix-its, checkup composition, and the remote-terminal request endpoint) returns "Free plan — technician services not included", and the command-dispatch path drops tech commands for free installs even if one is already queued. Remote commands were previously not license-gated at all — the Free plan is the deliberate exception, pinned by the release suite.
- Heartbeat enforcement: free installs get an empty add-on list, and the web-protection, advanced real-time, and email-protection policy flags are forced to 0 (real-time file protection is part of Free and is not touched).
- Upgrade converts the same key: a paid checkout carrying the account's Free key converts that license row in place (device stays registered; full entitlements arrive at the next heartbeat). Reseller commissions never accrue on free licenses (they accrue on the paid upgrade, like any paid sale). The 7-day trial is unchanged and coexists as its own license.
- Site & docs: /subscribe and the pricing cards gain the Free card first ($0, 1 device, truthful feature list, Get Free CTA into My Account); the homepage "Everything included" marks paid-only items and its intro no longer claims every plan is the full product; FAQ gains five Free Q&As (EN + ES, feeding the FAQ JSON-LD from the same list); the Terms gain a plain §2b Free plan section; the admin Licenses tab gains a plan filter including Free, the fleet list badges FREE devices, and the dashboard counts Free licenses; the admin guide documents the plan end to end.
- Custom rules no longer truncate (Workstream B): the client loader cap rises from 256 to 1,024 rules (Client 2.5.149), and the admin Detection rules surface now shows the published count with a visible warning when it exceeds the client cap — the cap lives in one shared constant on the server and is pinned against the client source by the suite. A fresh server seeds zero detection rules (verified on a clean boot); rules accumulate only from feed ingestion and admin authoring, and anywhere a database has grown past 256 the fleet was silently dropping the newest rules until this release.
Version 2.5.149
The Windows app knows the Free plan — and loads every detection rule.
- A Free device shows a calm "You're on the Free plan" banner with an Upgrade button — never an expired banner, never the daily payment popup (Free is a real plan, not a lapsed one).
- Paid-only toggles (web protection, advanced real-time protection) answer with a plain "Paid plan" message on Free devices instead of failing silently; the scheduled-scan settings show the same note, and scheduled scans stay off while free (manual scans are unaffected).
- On-device enforcement mirrors the server: the anti-ransomware/rollback module, the email protection module, and scheduled scans all report disabled while the tier is
free, and the client forces the paid-only modules off whenever a registration or heartbeat reports the free tier. - Free installs send their install credentials with the update check so the server can answer with the daily signature snapshot; paid installs keep the anonymous live manifest exactly as before.
- CRULE_MAX_RULES 256 → 1,024 in the custom-rules loader: the newest published rules are no longer silently dropped.
The server half of Batch 11 release C (the client half is Client 2.5.122, next section): fleet policy gains appfw_mode + allowed_apps + metered_mode (admin Fleet UI, sanitization, heartbeat delivery as policy_appfw_mode / policy_allowed_apps / policy_metered_mode); detection reports accept reason_texts, stored on a new detections.reason_text column and shown in the admin Detections list as a Why: line; the /download page explains saving the update files for offline import; the admin guide documents the new fleet fields. Suite: v25122-batch11r3-test.js.
Version 2.5.193
Customer reviews of technician checkups — shown only with the customer's permission AND the shop's approval. (Server-only release; the Windows client stays 2.5.148, Android 1.34.0, Linux 1.27.0, Term Helper 1.0.3 — no client or update-channel change.)
- The prompt: when a checkup request (2.5.191) reaches Done — by a clean diagnostic result landing or by Mark handled — My Account → Billing → Checks by your technician offers "How did we do?" for that checkup: a short plain-text review (600 characters), a display name defaulting to first name + last initial (customer-editable), and a permission checkbox, "You may show this review on aeantivirus.com." One review per completed checkup; the customer can edit or delete it while it is pending, and editing an approved review returns it to pending for re-approval. (Rejected reviews can be deleted but not edited — the shop's no is final for that text.)
- Permission is the hard rule: a review is displayed only when BOTH are true — the customer ticked permission AND an admin approved it. The admin approve endpoint refuses (400) when permission is missing, the public read filters on both conditions, and reviews without permission are private feedback: readable in admin, never rendered on any public page or in any JSON a public page consumes. Admin approval can never substitute for permission.
- Admin: a Customer reviews card on the Tech tools tab, next to Checkup requests — pending queue plus recently decided; private reviews are labeled "Private feedback — cannot display" with no Approve action offered (and the server would refuse it anyway); Reject hides permanently (record kept); a shop notification fires on each submission; submit / edit / delete / approve / reject are audit-logged (
review.*). No star ratings — deliberately: the account UI has no customer-facing rating pattern (the 2.5.130 health ratings are an internal admin metric, a different thing), so reviews are text-only. - Display: /compare (English + Spanish chrome) gains a "What customers say" section ONLY when at least one approved + permitted review exists — with zero, the section is completely absent from the served HTML (no heading, no placeholder). Up to 3 most recent, shown as text, display name, and month/year; the review words are the customer's own, untranslated, and rendered escaped everywhere (the release suite posts an HTML/script payload end to end and asserts it renders as inert text on /compare, in admin, and in My Account).
- No fabricated proof, ever: nothing is seeded — production starts with zero reviews and the section appears only when a real customer writes one, permits it, and the shop approves it.
- Copy surfaces: the Privacy policy gains a Checkup reviews entry (what is stored, the permission + approval rule, edit/delete rights) and its date moves to October 19, 2026; Help & FAQ gains "Can I leave a review of my checkup?" in both languages; the Terms are unchanged (a review is voluntary feedback, not a term of service). The admin guide gains the matching operator note.
Version 2.5.192
The comparison proof round: build-your-own add-ons with a counted catalog, a public sample checkup report, and flat-renewal pricing in the open. (Server-only release; the Windows client stays 2.5.148, Android 1.34.0, Linux 1.27.0, Term Helper 1.0.3 — no client or update-channel change.)
- Build your own protection: /compare (English + Spanish) gains a section contrasting the big names' fixed tiers — pay for the whole tier to reach one feature — with AEAntivirus's base protection plus add-on catalog. The printed count, 228 add-ons (128 paid, 100 free) — 227/127 when this section shipped in 2.5.192; the 2.5.214 Router Guard add-on joined the catalog — is computed from the live
ADDON_CATALOG + FREE_ADDON_CATALOG in the server code, and the release suite recomputes it from the code and asserts the page matches — the number cannot drift. - Sample checkup report: a public /sample-report page (English + Spanish) renders the real printable device-report format with clearly-labeled sample data — a prominent SAMPLE banner, a fictional device ("Sample — Smith Family PC"), and results whose check names are the real tech-tools catalog labels, with English summaries produced by the same
diagSummaryText function the real report uses. Linked from /compare's How it works ("See what your technician sees") and in the sitemap. Real reports stay behind owner/manager auth; the sample is marketing, never a customer record. - Pricing honesty, stated flatly: /compare gains a What you pay row — Custom at $4.99 per device per month with the automatic volume ladder (11–24 devices $4.49, 25–49 $3.99, 50+ $3.49) — and the owner's confirmed policy in plain words: your price is the price that renews; there is no first-year price that jumps. Competitor cells state only the general practice (first-year discount pricing that rises at renewal) — no named vendor's price is quoted. The same promise is added to the Terms (§3, Renewal pricing) and to Help & FAQ in both languages ("Does my price go up when my plan renews?" — no).
- FAQ structured data: the Help page now emits FAQPage JSON-LD generated from the same Q&A data the page renders (one source for both languages), so search engines see exactly what visitors see; the release suite parses the JSON-LD and asserts every question matches a visible question on the page. /compare also cross-links the Threat Guide (/threat-definitions).
Version 2.5.191
The comparison expansion, the switching guide's missing steps, and customer-requested checkups. (Server-only release; the Windows client stays 2.5.148, Android 1.34.0, Linux 1.27.0, Term Helper 1.0.3 — no client or update-channel change.)
- /compare grows: a How it works section (run a checkup → your technician reviews the findings and health trends → fixed remotely with your consent, or hardware recommended, quoted and approved before work, by walk-in or mail-in — priced by what the hardware costs, turnaround varies by repair, mail-in shipping paid both ways); a Remote help, on your terms section (per-session consent on the customer's device, 60-second auto-deny, the customer's own End session button, sessions recorded with transcripts the owner can review, fix actions never auto-billed, invoices draft-only); a three-way strip (antivirus only vs repair shop only vs AEAntivirus); and a local proof block for Mott Enterprises / Austin's Electronics (1381 Tomahawk Hwy Suite B, Harrells, NC 28444, 910-214-3670) linking to /about. English + Spanish, same safe-claims rules as 2.5.190.
- Printable one-pager: /compare/print renders a standalone print-ready sheet — headline, condensed comparison table, how-it-works steps, and the shop's name, address and phone — for the shop counter; /compare gains a Print button and a print stylesheet.
- /switch restructured (not duplicated): the 2.5.129 switching guide now runs in the order customers actually need — Step 1: turn off the old plan's auto-renew first (generic billing wording only; nothing promised about any maker's policies), Step 2: uninstall the old antivirus, Step 3: install AEAntivirus (license key only), Step 4: run your first Full checkup and let your technician review it. The page gains its Spanish edition, and Help & FAQ links it from a new "How do I switch from Norton or McAfee?" entry in both languages.
- Request a technician checkup: from My Account → Billing → Checks by your technician, a customer can request a checkup per device with an optional short note — one open request per device, cancellable while open, with status (Requested / In progress / Done) shown beside their checks. Requests are work items, never charges: requesting creates no fee, and the existing diagnostic-fee rule applies only when a check actually completes. Open requests surface on the admin Tech tools tab and the device page (plus a shop notification); queueing a check moves a request to In progress, a clean landed result completes it automatically, and Mark handled closes one without a run. Create / cancel / complete are audit-logged.
Version 2.5.190
A public competitor comparison page: /compare sets AEAntivirus beside Norton 360 + Premium Services, McAfee + its Virus Removal Service, Trend Micro Maximum Security, and AVG + AVG Go / Premium Tech. (Server-only release; the Windows client stays 2.5.148, Android 1.34.0, Linux 1.27.0, Term Helper 1.0.3 — no client or update-channel change.)
- The comparison page: a new public page at /compare (English + Spanish, following the /tech-services view pattern) with the full comparison table from the approved draft: how the technician is sold (built into AEAntivirus; a separate service from the big names), what the technician does, the 49 built-in health checks (Windows; 19 Linux, 8 Android), the technician seeing findings before the customer has to explain anything, the customer's own account view of checks and repairs, the remote-session consent protections, hardware help, and who actually helps you (the Mott Enterprises / Austin's Electronics team in Harrells, NC).
- Safe claims only: the page never calls AEAntivirus the only antivirus with a technician, never claims round-the-clock technician access for AEAntivirus, never claims better detection, and never prints a flat hardware price, a promised turnaround, or an invented repair promise. Hardware copy states the shop's terms exactly: priced by what the hardware costs, quoted and approved before work, turnaround varies by repair, mail-in customers pay shipping both ways. A new release suite pins every forbidden claim absent.
- Site wiring: the homepage gains a comparison teaser under the headline "Antivirus, plus a technician. All in one place." plus a card in Everything included → Technician services; /compare joins the footer Support column and the sitemap; the Tech services page links to it; Help & FAQ gains two entries in English and Spanish — "How is AEAntivirus different from Norton or McAfee?" and "Do you offer technician help?"; and the admin guide gains a matching note. No Terms, pricing, or feature-scope change — nothing about the product's terms moved.
Version 2.5.189
The branding-consistency release: one brand everywhere. An About page for the app, the wordmark as the only brand art on every surface, every icon redrawn from it, and simpler wording across the whole family. The Windows client is 2.5.148 (below); the Term Helper is 1.0.3; Android is 1.34.0 (below); Linux and the extension source are unchanged (the extension's icons are redrawn too, for its next build).
- About, in the app: the Windows client gains an About page (sidebar -> Help -> About) with the wordmark, the version, and the same About-us story the website tells — Mott Enterprises, Austin's Electronics in Harrells, NC, what the shop does, and how to reach us — plus a button that opens aeantivirus.com. The Term Helper gains the same facts in its own new About dialog (helper 1.0.3, below), and the Android app tells the story in its What's New.
- One brand art: the old logo image is gone from the Windows client's window, the Term Helper's sign-in window, the website header and footer, and the Android app's screens. The AE Antivirus wordmark is the only brand art anywhere now, and every square icon — the Windows app, the installer, the helper, the website favicon and touch icons, the Android launcher, the extension — is the block AE letters cropped from that same wordmark.
- Simpler wording: the old name for the protection modules is gone from product copy across the family — the apps, the site, the admin guide (older entries included), emails, and this changelog. They are simply "protection" now (or their own module names: Real-time protection, Web protection, and so on). Only internal keys and field names keep the old spelling, for compatibility with installed devices — no setting, behavior, or protection logic changed anywhere.
- Helper 1.0.3: a proper About dialog (an About button on the sign-in window: the wordmark, "AE Term Helper", its version, and the same About-us facts as the client's About page), the new AE-letters icon, and the in-window logo removed (its name and version move left under the wordmark strip). Every earlier safety pin holds: HKCU-only aeterm:// registration, loopback-only listeners, no process launching, no shell-outs, tokens and passwords memory-only.
Version 2.5.148
A new About page, the wordmark alone at the top of the window, and the new AE-letters icon.
- About page: a new About entry under Help in the sidebar opens a page with the AE Antivirus wordmark, the version, and our story — AEAntivirus is designed and coded in the USA by Mott Enterprises, the company behind Austin's Electronics in Harrells, North Carolina (PC and phone repair, virus removal, and rural internet service). The shop's address (1381 Tomahawk Hwy Suite B, Harrells, NC 28444) and phone (910-214-3670) are right on the page, with a "Visit aeantivirus.com" button.
- New look: the old logo image that sat beside the wordmark in the sidebar head is gone; the wordmark heads the sidebar on its own. The app's file, taskbar and tray icons are now the block AE letters cropped from the wordmark. Status wording is simpler too ("All protection running normally"). Nothing else moved — every page and setting is exactly where it was, and protection is unchanged.
Batch 11 release B (server-only; the Windows client stays 2.5.121):
- Detection triage board. New side table
detection_triage (detection_id PK, state new|investigating|resolved, note, updated_by, updated_at) — deliberately a side table: a detection with NO row is state new, so there is no migration flood, and triage stays admin-only data (it is selected only by the admin list query; a test asserts the note text and field names never appear in /account or the data export). The admin Detections list now joins triage state/note/by/at, accepts a ?state= filter, and returns per-state triage_counts computed over the whole unresolved board. The UI gains Triage filter chips with counts, a per-row state select + note editor, and a bulk Mark selected resolved (triage). New POST /api/admin/detections/triage (owner) writes one row or a bulk state change. Interplay: both false-positive flows — the admin verdict endpoint AND the client verdict endpoint — call the shared triageSetFp() helper, writing a resolved row with the note "Marked false positive"; the confirmed-threat flow never touches triage. Triage is workflow state only: it does not resolve detections, credit ML signals, or alter verdicts. - Reseller custom domains (server half). New
reseller_domains table (reseller_id, domain UNIQUE, verified_at). Pure validation + the DNS rule live in server/lib/reseller-domains.js: hostnames are validated (2+ labels, no IPs, not this site's own host, no localhost/.local), and verification succeeds ONLY when a server-side CNAME lookup returns the site's configured public host (PUBLIC_URL hostname) exactly (case/trailing-dot-insensitive). Reseller portal gains a Custom domains card: add a domain, the DNS checklist (CNAME to the public host + the Render dashboard step for TLS), and a Check button that runs the lookup and stamps verified_at only on a match. Serving: brandForHost(req) hooks in FRONT of brandForAccount in render() — a verified domain in the Host header resolves that reseller's brand (their name, through the same brand object the page machinery already consumes); unknown/unverified hosts return null and fall through to current behavior byte-for-byte. Admin: GET /api/admin/reseller-domains + owner-only removal, with a Reseller custom domains card in Customers → Resellers. Austin's remaining step per domain (also documented in the portal card and admin guide): add the domain under Custom Domains in the Render dashboard so requests reach the service with a valid TLS certificate — server verification only proves the DNS points here. - Docs: admin guide gains triage + reseller-domain entries (including the Render step); homepage "New in" rotates (2.5.145 folds). No ToS change: reseller branding was already described in the reseller program terms, and triage is internal admin workflow.
- Suite:
v25148-batch11r2-test.js (56 checks: triage transitions/counts/filters/bulk/note-only/invalid state, FP interplay through BOTH verdict flows, TP non-interference, customer-surface leak checks; domain hostname/CNAME unit tests with a stubbed resolver, claim/duplicate/check-refusal black-box, Host-header brand serving verified vs unverified vs unknown, admin list + removal + re-claim).
Version 2.5.188
The AE Antivirus wordmark arrives in the Windows app and the Term Helper — pure branding, no behavior change. The Windows client is 2.5.147 (below); the helper is 1.0.2; Android, Linux and the extension are unchanged.
- The wordmark, in both programs: the glowing blue DNA "AE Antivirus" lockup — the same art the website header and the Android app already carry — now heads the Windows client's main window and the AE Term Helper's sign-in screen. In the helper it sits as a brand strip above the logo, with "Term Helper" and "Version 1.0.2" kept directly under it so the helper's own name stays unambiguous.
- Helper 1.0.2: the helper exe (server/assets, owner-only download at /admin/term-helper.exe, never in the client update channel) embeds the wordmark as a resource and stamps 1.0.2 in its version resource, title and About line. Every 1.0.1 safety pin holds: HKCU-only aeterm:// registration, loopback-only listeners, no process launching, tokens and passwords memory-only.
Version 2.5.147
The app window wears the full AE Antivirus wordmark.
- Wordmark header: the DNA wordmark now appears in the sidebar head of the main window, beside the logo, exactly where the text name stood — the app opens to the Status page under the full brand lockup. (The loader for this art has been in the client since 2.5.3; the embedded resource itself was missing, so the header silently fell back to drawn text. The resource is now actually embedded, and a test pins its bytes inside both shipped exes.)
- Nothing else moved: the logo, the version line under the name, every sidebar entry and every page are exactly where they were — this release changes no behavior, no setting and no protection logic.
Batch 11 release A (server-only; the Windows client stays 2.5.121):
- Self-service password change (closes the Batch 10 gap). New
POST /api/account/change-password: verifies the current password with the existing scrypt check, enforces the SAME policy as registration on the new password (8+ characters + the breached-password/HIBP check, fail-open on HIBP outage as at registration), and requires a valid authenticator code when the account has 2FA enabled. On success the hash is updated, other sessions are revoked — ALL sessions including the current one when "sign out everywhere" is ticked, otherwise all except the current one — a security email goes out through sendMail (kind security, so it appears in the Email log), and the change is audit-logged (account.password.change). The form lives in My Account → Security. Scope is the signed-in change only: there is still no forgot-password flow (verified — none existed; the Help FAQ previously claimed a reset link on the login page, which was wrong and is corrected in this release), and a customer who cannot sign in at all still needs a support conversation. The stale 2.5.141 copy is fixed everywhere it lived: the Help sign-in-history answer and the admin guide now point at the new flow instead of "contact support so we can reset your password"; the sign-in-history revoke-all behavior itself is unchanged. - Outbound webhooks for Business / Enterprise / Custom plans (same
accountApiEligible gate as customer API keys). New customer_webhooks + customer_webhook_deliveries tables. The signing secret (whsec_…) is shown once at creation; it is stored as issued because it must be readable to compute HMACs at delivery time (the same treatment as other server-side retrievable secrets, e.g. the backup passphrase — API keys, by contrast, are verified by hash and stay hashed). Events: detection.confirmed (fires from the same confirmed-verdict path as the threat-alert emails — heuristic suspicion and FP-excluded paths never fire), license.expiring (fires from the expiring-soon reminder send site, the one place the expiry-warning email is sent per license), and device.offline (a sweep on the alert-rules 15-minute cadence fires once per episode when a device crosses 24h without a heartbeat — the admin offline rules' default window; the episode row clears on the next heartbeat). Delivery (pure core in server/lib/webhooks.js): HTTPS POST of a JSON envelope with X-AE-Signature: sha256=<HMAC> over the raw body, 5-second timeout, no redirects, 1 attempt + 3 retries with backoff, EVERY attempt journaled. SSRF discipline: registration refuses non-HTTPS URLs, URLs with credentials, localhost, and private/loopback/link-local literals; delivery re-resolves and refuses when ANY resolved address is private, and the connection is pinned to the validated address so DNS rebinding can't swap it. /account → Security gains the Webhooks card (add with URL + event checkboxes, list with last-delivery status, delete, delivery log, Send test ping). Locked-out caveat for tests: the delivery engine's resolver/transport are dependency-injected (lib/upgrades-util.js pattern), so the suite stubs them — the guard itself is unit-tested separately and is never bypassed in production code. - Monthly fleet report email. New
monthly-fleet-report job (checked every 15 minutes; acts on the 1st of the month for the PREVIOUS month, with a state file so restarts never double-send). Population: Business/Enterprise/Custom owners with at least one install who haven't opted out (new accounts.fleet_report_optout, toggle in My Account → Security, default ON for the tier). Every figure comes from real records: protected device count, detections blocked in the month (exact window), top 5 threat names, devices on outdated client versions (vs the manifests on disk), and licenses expiring in the next 30 days. Sent through the email-template system (new fleet_report template key, admin-editable) via sendMail (kind fleet-report). The job is listed in the jobs viewer as NOT Run-Now-safe (it sends email); the owner trigger is POST /api/admin/fleet-report/run (audited), same pattern as the win-back drip. - Docs: Help/FAQ gains password-change, webhook, and fleet-report answers (and the forgot-password answer is corrected); Terms §3b/§3c and the Privacy policy now describe webhook event sharing and the fleet report; the admin guide documents all three features; homepage "New in" rotates (2.5.144 folds) and "Everything included" gains cards for webhooks, the fleet report, and password change.
- Suite:
v25147-batch11r1-test.js (98 checks: password-change semantics incl. 2FA + both revocation modes + mail_log; webhook SSRF/delivery unit tests incl. in-test signature verification and retry counts; black-box tier/registration refusals, all three event fire points; fleet-report exact figures, tier/zero-device/opt-out exclusions, jobs-viewer state).
Version 2.5.187
The AE Term Helper grows up: AE branding, the client's styling, and one-click opening from the admin panel (helper 1.0.1) — plus a signature-channel coherence fix. The Windows client is unchanged (stays 2.5.146; manifests and installers untouched).
- Helper 1.0.1 — branding: the AE logo is the helper's file and window icon (the same icon asset the client and installer carry, embedded as a resource) and sits beside the sign-in form, with the product name and a "Version 1.0.1" line presented the way the client presents its own name and version. The exe's version resource and About line read 1.0.1.
- Helper 1.0.1 — styling parity: a pass against the Windows client's current values — the ui_theme.h palette (near-black blue background, panel tone, light text, the 0,229,255 cyan), Segoe UI at the client's 10pt text / 12pt bold heading sizes with its grayscale antialiasing (ClearType fringes on the dark theme), and the client's unified button treatment (dark vertical gradient, radius-14 corners, an always-visible 2px cyan border that brightens on hover/focus, a soft outer glow, focus rect). The 1.0.0 safety pins all hold: loopback-only listeners, no process launching, tokens memory-only, password zeroed after the sign-in request is built, the ini still persists only the remembered email.
- Open the helper from the panel (aeterm://): the Remote terminal card gains an Open in Term Helper button, an Open this session in Term Helper link on a live session plus the same link on each session-history row, and a Download the helper link for a new admin PC. On first run the helper registers the
aeterm URL protocol under the current Windows user only (HKCU\Software\Classes\aeterm — no admin rights, nothing machine-wide) and re-registers if the exe moves; that registration is the helper's only registry write. A session link carries the session id only — never a token or credential — and is never authentication: the helper still requires its own sign-in, after which it lands on the sessions list with that session highlighted (it never opens a tunnel by itself). A second launch forwards its intent to the already-running copy (named mutex + WM_COPYDATA) and exits. The browser's first-click "Open AE Term Helper?" prompt is expected and documented in the admin guide. - Signature channel coherence: the signature publish path (publishSignatures) now writes signatures.db and each platform manifest's sigs fields (sigs_version + sigs_sha256) atomically — temp file + rename — inside the same publish cycle, so a publish can never leave the static channel torn or half-updated: after any publish, every platform manifest records the served database's actual hash at one shared sigs_version, which is exactly what the 2.5.160 update-channel drift canary verifies.
Version 2.5.186
A removed device can be restored by the shop owner, and the admin console shows a loading state instead of a blank page.
- Restore removed device (owner only): removing a device from a license tombstones it — it cannot check in or re-register with that key, and for customers that stays permanent (the anti-theft point). But the product shipped with no owner override at all: a device removed by mistake — the motivating case was the owner's own bench PC — could never come back. Fleet → Installs now has a Removed devices card listing every tombstoned install (hostname, masked license key, device-fingerprint prefix, reason, removal date) with a Restore button. Restoring deletes the install's tombstone row plus its (license key, device fingerprint) pair — the two keys registration refuses on — is audit-logged, and the device can then re-register with its original key at its next attempt. Managers and support get 403 on both endpoints; there is no customer surface. The My Account removal warning now says the device cannot re-register "unless the shop restores it" (EN + ES).
- Console loading state: the admin panel renders a "Loading the console…" card from first paint; an independent boot-watch script replaces it with a plain slow-boot notice ("Still loading — this connection or device is responding slowly" + a Reload button) if the console's main script has not booted within about 8 seconds. A slow or blocked script can no longer present as a silent black page (the 2026-10-18 blank-panel incident).
Version 2.5.146
Startup can never be taken down by one bad quarantine file again — and a hard crash now leaves evidence.
- Quarantine sweep hardening: the startup encryption sweep gives each raw quarantine file one attempt. A file that fails processing is skipped, flagged
<name>.quar.bad (out of the sweep's scan set), logged with the reason, and never retried; the sweep continues with the next file and returns normally to startup. Files past the encrypt path's 256 MB cap are flagged immediately on the same rule. (The 2.5.141 crash loop was one raw .quar retried on every launch.) - Encrypt-at-write quarantine (the stored-raw closure): every quarantine write path funnels into one writer, and it now writes the stored copy already encrypted — AEAVENC1 frame via foldenc, verified by decrypting it back before the original is removed — instead of moving the file raw and encrypting as a best-effort afterthought. The audit found exactly one raw-store route: the writer's post-move encrypt step failing (key unavailable) and keeping the raw file with only a warning. If encryption is unavailable the file is still quarantined, the raw store is logged at error level, and the hardened sweep flags it rather than retrying it forever. A newly quarantined file on disk begins with the AEAVENC1 magic bytes.
- Own tools are never quarantined: AE-Term-Helper.exe joins AEAntivirus.exe and afunix.sys in the never-quarantine set. The standalone admin helper was not an installed component, so the existing own-files protection never covered it — and the client quarantined its own helper the moment it was downloaded.
- Startup step logging: the engine logs an enter/done line for each step in the window between reputation prefetch and the watchers (canary planting, quarantine sweep, missed-scan catch-up, watcher start), so any future startup death names its exact step in a single launch.
- Watchdog crash evidence: a fail-fast crash (0xC0000374) bypasses the in-process crash reporter entirely — no minidump, no pending report. The watchdog process now writes a pending crash report (kind: watchdog) carrying the last 32 KB of the client log when it observes that exit code, or 3+ exits within 10 minutes, throttled to once per hour and written only after the relaunch (or on the backoff path) so reporting never delays a restart. The engine uploads it through the existing crash-report channel at its next start.
- Scope correction: the My Account reset now resets only that device's local counters (Austin 2026-10-16: "The reset in the my account should only reset that devices local counters"). In 2.5.144 all three reset doors did the full thing — server mirror zeroed + server-side detection records deleted immediately, plus the
reset-stats command. The My Account door is now queue-only: it queues reset-stats for the device (which zeroes its own counters + local history on check-in, exactly like the in-app button), it never deletes the detection history stored in the account, and it never zeroes the server mirror directly — the mirror keeps showing the device's real totals until the device executes the reset and heartbeats its zeroed counters, then drops to 0 on its own. The race guard is decoupled to match: the 2.5.144 heartbeat substitution (hold the stored lifetime at 0 while a reset is unacknowledged) was only correct because the server had pre-zeroed, so it now arms only for guard-armed resets. The distinction rides on a new device_commands.origin column ('admin' | 'account'; rows queued by 2.5.144 carry '' and keep their original admin semantics, since their server side *was* zeroed). If both doors fire for one device, the admin semantics win: an admin reset upgrades a pending account-queued command in place (origin → 'admin', guard arms, server zeroes + deletes) instead of queueing a duplicate, and a customer click while an admin reset is pending queues nothing new. Unchanged: the admin install-detail reset (full behavior, owner-only) and the in-app Reset statistics button (already purely local) — no client release. Copy follows the code: the My Account confirm dialog and the Help FAQ now say the account reset resets the counters on the PC itself and keeps the account's stored history; the admin guide notes the account door is queue-only. Account resets stay audit-logged (install.reset_stats, with the queue-only nature in the detail). Suite: v25144-reset-stats-test.js extended (account door: records intact, mirror untouched, no substitution while pending, mirror follows after ack + zeroed heartbeat; mixed account→admin case pinned; admin door's full sequence unchanged). - Server-only release; the Windows client stays 2.5.121 and Linux/Android are untouched.
Version 2.5.185
Tech tools now cover Linux and Android, with honest per-platform check sets — the Windows client is unchanged (stays 2.5.145; Windows manifests and installers untouched). One server-side matrix is the single source of truth: Windows 49 checks, Linux 19, Android 8.
- The matrix (server-side, one table):
DIAG_PLATFORM_TOOLS maps each platform to the checks it can honestly run. Linux: health, spec, hardware, drive, ssd, space, battery, net, time, temp, services, users, programs, events, crashes, startup, adapters, posture, bootperf. Android: battery, space, net, spec, health, programs, posture, time. macOS is recognized but has no checks yet — every request for one is refused honestly instead of queued. - Enforced everywhere, with the reason: queueing an unsupported check for a device's platform is refused (single queue, bulk queue — where it lands in the per-device skipped list with the reason — schedules at creation and at fire time, and the device ingest endpoints themselves). The refusal reads "Hardware inventory is not available on Android devices" — never a fake queue, never a silent drop.
- Per-platform checkups: the Full checkup runs each device's own set — Windows its established 8, Linux 6 (health, spec, drive:all, ssd, space, posture, bootperf), Android its 8 — and the panel labels exactly what will run per platform. Multi-device runs apply per-device eligibility and report skipped devices with reasons.
- The panel follows the device: the Tech tools device panel and card offer only that device's platform tools, the diag API returns the platform, its supported set, checkup composition and the full matrix, and the customer's Diagnostics & repairs rows name the device platform ("Office PC (Linux)").
- Findings audited per rule: every fleet-findings rule keys on fields whose meaning the Linux/Android collectors preserve (a drive is only "bad" on a real SMART FAILED; battery/SSD wear fires only from real reported values). A Linux desktop with no battery answers exactly like a Windows desktop ("battery: none") and bills the same; error-shaped no-data results never bill, on any platform.
- Copy: the admin guide gains the platform matrix with every skipped tool's reason; /tech-services (EN + ES) and the homepage state the three counts (49 / 19 / 8); Terms §4s and the Privacy diagnostics item name Android and Linux collection in the same plain style.
Version 2.5.184
Raw tunnels + the AE Term Helper — the terminal feature is complete: inside a live, consented session the technician can now reach the device's OWN Remote Desktop (3389), VNC (5900), and SSH (22) services, through a native helper app on the technician's own PC (AE Term Helper 1.0.0, x64). Web tunnels are unchanged.
- Raw tunnels (kind "raw"): the owner opens one from the live session view in the admin panel or from the helper — same endpoint, owner-gated either way. Target ports are exactly {22, 3389, 5900}, enforced by BOTH validators (the server's raw validator and the web validator are disjoint: a web tunnel to a raw port is still refused, and a raw tunnel to a web port is refused), and the device connects to the literal address 127.0.0.1 only. Raw streams exist only inside an active consented session and die with it.
- The relay: the device speaks its existing HTTPS chunk transport (open/data/close ride the ~1.2 s term poll; device output posts back on
/api/term/raw-output / /api/term/raw-end); the helper speaks HTTPS chunk posts + drain-polls under /api/term/relay/* — the same pattern, no new transport. Chunks are capped at 32 KB decoded bytes per post; byte counts accumulate per stream and per tunnel (decoded bytes). - Per-tunnel relay tokens: opening a raw tunnel issues a single-tunnel token (
rt- + 24 random bytes) that expires no later than the session's own 2-hour cap and is tombstoned the moment the tunnel closes or the session ends — later relay calls get 410 (unknown token: 401; a stream from another tunnel: 404). The token reaches the helper in the tunnel-open response and, for a tunnel the panel opened, through the owner-only session view — both consumers are the owner. - Helper sign-in: the helper signs in AS the admin —
POST /api/admin/term/helper-login mirrors the panel login exactly (same lockout bucket, same failure logging, same TOTP rules, same "enroll 2FA first" gate) and then refuses non-owner roles with 403: this whole family is owner-only. It returns a helper token (ht-, 8 hours, memory only) accepted as a Bearer credential on the tunnel-open/close, helper-sessions, and helper-logout endpoints. Tokens for the panel's cookie sessions are untouched. - Transcripts stay metadata-only: raw bytes (SSH/RDP/VNC traffic) are NEVER stored server-side — the queues are transient in memory. The transcript records note entries only: the tunnel opening (port + service), each stream's close (reason + bytes relayed), and the tunnel close (total bytes). The audit
term.tunnel_open / term.tunnel_close payloads gain kind, port, and bytes. Raw data never appears on any customer-facing API. - The helper app: source in
src/term-helper/ (C + Win32 + WinHTTP, built by build.sh with the same MinGW toolchain). Sign-in (server URL prefilled, email, password, authenticator-code field enabled when the server says TOTP is required), the owner's live sessions, open a tunnel (service + editable local port, defaults 13389/15900/12222), attach to a raw tunnel the panel already opened, multiple simultaneous tunnels, per-tunnel byte counters, and the exact connection string as copyable plain text (e.g. mstsc /v:127.0.0.1:13389). Its local listener binds 127.0.0.1 ONLY — one relay stream per accepted connection. - Credential hygiene (source-pinned): the password is zeroed right after the sign-in request is built and is NEVER written to disk or registry; the only persisted value is the email (checkbox, in an ini next to the exe); relay and helper tokens live in memory only; sign-out stops the listeners, closes helper-opened tunnels server-side, and clears the tokens.
- Delivery: the built helper ships at
server/assets/ae-term-helper.exe and is served ONLY from the new owner-gated route GET /admin/term-helper.exe (manager/support 403). It is NOT in the client update channel — the installer, updates tree, and the update manifest carry no trace of it (the suite pins that the manifest JSON contains no "helper"). - Copy: the admin guide gains the raw-tunnels + helper section; Terms §4u and the Privacy terminal entry now describe raw forwards to the device's own RDP/VNC/SSH services inside a consented session, content never recorded (metadata only); the /tech-services terminal paragraphs (EN + ES) gain the matching sentence. Soak endpoint count: 187 admin GET (+ helper-sessions; the helper download route sits outside the soak's API count).
Version 2.5.145
Raw streams inside terminal sessions — the device side of the raw tunnels above (needs Server 2.5.184 for the relay).
- What the client does: inside a live consented session, a raw open request names a stream id and one of the fixed ports {22, 3389, 5900} (validated on the device by
term_raw_port_ok — anything else is ignored). The client opens a TCP connection to literally 127.0.0.1 on that port — the address is a constant in the code, never transported — and shuttles byte chunks both ways over the existing terminal chunk transport, base64-encoded, per-stream sequenced, capped at 32 KB per chunk, with a 1 MB per-stream backlog cap. - Honest failures: a refused connection on the device (the service is not listening) is reported back as
connect-refused and the stream closes — the technician sees the real reason, never a fake success. Streams are capped at 8 per session, and EVERY stream closes when the session ends, whatever the end reason. - Unchanged: web tunnels work exactly as before; the consent dialog, the End session button, the idle/max limits, and the constant cmd.exe shell path are all as shipped in 2.5.144. No new client command type was added — raw streams ride the existing
term-open session.
- The About page grows up (Austin 2026-10-16: "Add more to the about us"). Six new sections join the existing page — nothing already there was removed. Our story: AEAntivirus began in the repair shop — Austin's Electronics in Harrells, NC kept seeing the same infections come back through the door, so the same hands that fix infected PCs built the protection those customers should have had. What we stand for: plain-English alerts, no scare tactics (the fake "your PC is infected — call now" pop-up is scareware, and our own product blocks it), honest "AI" language (a trained model is called a model; signatures and heuristics are called what they are), and fair pricing. What we make: cards for Windows, Linux, and Android, Family Safety, AE for Business, and the reseller program, each linking to its page. How support works: contact reaches the people who built it, via the support inbox, Help & FAQ, and the user guide. Local roots: Mott Enterprises — the umbrella behind AEAntivirus and Austin's Electronics — owner-run, rooted in Harrells, NC, serving its own community first. The pre-existing timeline card is retitled "How we got here" so the page doesn't carry two "Our story" headings.
- By the numbers — counted, never typed. The new stats row (threat types in the Threat Guide, releases in the public changelog) is computed at request time from the sources of truth themselves: the guide count is exported from the Threat Guide's own data (
guideCount in views/definitions.js), and the release count is the parsed changelog's length — the same data /changelog renders. No customer counts, founding claims, or estimates were added; numbers that couldn't be sourced from the product were left out. - Server-only release; the Windows client stays 2.5.121 and Linux/Android are untouched. Suite:
v25145-about-test.js (version plumbing computed from package.json, all six sections over real HTTP, both rendered numbers asserted against independently computed sources, preserved-content and no-invented-claims guards).
Version 2.5.183
Terminal session billing — an ended session can be attached to a repair charge (server-only; client stays 2.5.144, installers and manifests untouched).
- Never automatic: a terminal session ending creates no fee, and there is no per-minute rate anywhere in the software. Billing happens only when the owner explicitly attaches an ended session to a repair charge from the Tech tools panel's session history.
- The attach: the dialog prefills the description —
Remote terminal session — <device label> — <duration as h:mm> — <date> (editable, at most 500 characters) — with labor and parts starting empty. The amount goes through the exact Record-a-repair path (createServiceFee, kind repair): whole cents, total must be over $0, the owner types the amount — nothing is pre-priced. The row lands in the same Service fees ledger and goes out on the same draft-only Stripe invoices as any other repair. - Linkage, never double-billed: the fee id is stamped on the session (
term_sessions.repair_id, NULL until billed) only after the repair row is created, behind a NULL guard — a second attach is refused server-side (409), and the panel entry then reads "Billed — repair #N ($X.XX)" with the action gone. If repair creation fails validation, no linkage is written and the error shows inline in the dialog. - Who can: owner only end-to-end — the attach endpoint sits behind the terminal owner gate (managers and support get 403), and sessions that were denied, timed out, or never answered cannot be billed: only a session that actually ran and ended can be attached.
- Customer view: the charge appears under My Account → Billing → Diagnostics & repairs with the session description, like every other repair. The attach is audit-logged (
service_fee.create, naming the session id and the fee id). The device's service notice is deliberately not turned on by an attach — the session is over, and the notice is for work in progress. - Copy: the admin guide's terminal section gains the billing paragraph; Terms §4u and the Privacy terminal entry each gain one sentence (a session may be billed as a repair when the shop records it as one, under the same terms as other repairs). The /tech-services terminal paragraphs (EN + ES) make no billing claim and are unchanged. Endpoint count: 186 admin GET (unchanged — the attach is one new owner-gated POST).
Version 2.5.182
Remote terminal sessions + device web tunnels in Tech tools — owner-only, consent-first, fully recorded. The admin Tech tools device panel gains a Remote terminal section (rendered for the owner role only; every endpoint below requires the owner role — managers and support get 403, stricter than the read-only checks):
- Request + device consent: "Request terminal session" queues the new
term-open:<session_id> command (session ids are server-generated, [A-Za-z0-9-], ≤64, validated by both validators and the client parser) and creates a term_sessions row in state requested. The customer sees a consent prompt on their own PC (see Client 2.5.144) and the panel shows the outcome live: waiting / denied / timed out / active with an elapsed timer. A second request while one is pending or active is refused (409) with the reason. - Session lifecycle + limits, enforced at both ends: 15-minute idle timeout (no input), 2-hour maximum, the session ends when the device stops polling (transport drop / restart) — the server sweeper ends stale sessions every 30 s and records the end reason honestly (
admin, customer, idle, max, disconnect, denied, consent-timeout, no-response, shell-exit). Requests nobody answers expire after 10 minutes. - Transcripts: every input line and output chunk is stored per session in
term_transcripts (output capped at 1 MB stored, with an explicit truncated flag and a note in the transcript when the cap is hit). The panel's session history opens any past transcript; transcripts are owner-only and are never exposed on the customer API. The store is deliberately in neither the prunable-stores list nor the retention-class spec — transcripts are never auto-purged (source-pinned). - Web tunnels: inside an active session the owner can open a forward to the device's OWN web interface —
http(s)://127.0.0.1:<port> with port in the fixed allowlist {80, 443, 8080, 8443}; no other host is representable in the protocol, and the client enforces the same allowlist and loopback target. The admin's browser reaches it through a session-bound proxy URL under /admin/term-tunnel/<session>/<tunnel>/… gated by the owner admin session and the live session — a dead session's URL dies with it (410). Tunnel bytes are counted (not stored); every open/close is audit-logged and noted in the transcript. Raw TCP forwards (RDP 3389 / VNC 5900 / SSH 22) are NOT in this version — the panel doesn't offer them and both validators refuse them: the host exposes no public TCP ingress and there is no admin-side helper yet; a future helper app could add them. - Transport: the rstream WebSocket relay is screen-frames and billing-shaped, so terminal traffic instead extends the diag/fix chunk-endpoint pattern symmetrically — device-authenticated endpoints under
/api/term/* (consent, poll, output, end, tunnel-response), never admin-cookie. The device polls about every 1.2 s while a session is live; input lines and pending tunnel requests ride the poll response, output chunks and tunnel responses post back. Line-based interaction is the v1 contract (full-screen console apps are not supported — stated in the panel and the admin guide). - Audit:
term.request / term.consent / term.denied / term.start / term.end (with reason) / term.tunnel_open / term.tunnel_close, all with the admin identity. New endpoints: 2 admin GET (term state per install, transcript) plus the device-auth /api/term/* set and the owner-gated tunnel proxy — 186 admin GET endpoints total. - Ride-along repair — parental-controls schedule editor: two regexes in the served /admin script rendered
\d as a literal d, so the schedule editor's per-day parser (/^([0-6]):(d{4})-(d{4})$/ as served) never matched a saved schedule like 1:0700-2100 — per-day start/end fields always loaded blank — and its time normalizer (/^(d{1,2}):?(d{2})$/) rejected every real entry, so saving a parental schedule with any day filled always failed with "enter HH:MM-HH:MM with start before end". Both are fixed at the source (\\d in the template literal) and the served bytes are pinned in the new suite. - Terms gain §4u "Remote terminal sessions" and Privacy a matching entry (consent per session, SYSTEM-level access stated plainly, transcripts kept with the service history and owner-only, end-any-time, tunnels limited to the device's own web interfaces); /tech-services (EN+ES) and the help FAQs (EN+ES) describe the terminal honestly, consent and recording included. New suite: v25182-term-test.js.
Version 2.5.144
The client half of remote terminal sessions (server 2.5.182). New src/term.c (+ term.h), with term-open:<session_id> wired in all four places (server VALID_CMDS + remoteCommandIsValid, client cmd_parse_one, both dispatchers):
- Consent in the customer's session: the service hands the request to the tray by file (the remote-start pattern); the tray shows a consent dialog modelled on the restart countdown — it names Austin's Electronics, says the technician can type commands and see the results on this PC, says the session runs with full system access, and says it is recorded and kept with the service history. Allow / Deny; 60 seconds with no answer is a deny. The device journals
term_consent_shown / term_consent_allowed / term_consent_denied / term_consent_timeout. A denied or timed-out session never spawns anything; in tray-only mode the command defers un-acked to the service (the remote-start discipline). - The shell: on consent the service spawns the FIXED
%SystemRoot%\System32\cmd.exe only (the path is a constant built from the system directory — nothing in any message can name a program; source-scan pinned) in the SYSTEM context on its own thread, with ConPTY (CreatePseudoConsole, loaded dynamically) on Windows 10 1809+ and plain pipes as the fallback. Admin input is written verbatim; console output is converted from the OEM code page and posted back in chunks. An idle session never blocks diag/fix work. - End it any time + limits on-device: while a session is active the tray window shows a persistent banner — "A remote terminal session is active on this PC" — with an End session button that kills the shell and records the end as the customer's (
term_ended_by_customer). The client also enforces the 2-hour maximum itself even if the server is unreachable, ends the session when the shell exits, and reports a transport drop honestly after 90 s without a successful poll. - Tunnels: within a session the client fetches tunnel requests from the device's own loopback only — the target host is the literal
127.0.0.1 in the code, the scheme is http or https, and the port must be in {80, 443, 8080, 8443} (the pure seams term_tunnel_port_ok / term_tunnel_scheme_ok are unit-tested headlessly, together with the session-id validator, the consent state machine, base64, and JSON escape/unescape in the new term-test.c). - The client's What's New page documents terminal sessions, the consent prompt, the recording, and the End session control. The completion-wording rule holds: "safe to restart" appears nowhere in the sources or the shipped binaries (guard-pinned).
- Reset counters, for admins and customers (Austin 2026-10-16). The install detail view gains a Reset counters tile, and every device row on My Account gains a Reset counters button: the device's lifetime threat total is zeroed and its server-side detection records are deleted immediately, and a new
reset-stats remote command is queued so the PC zeroes its own counters when it checks in. *(Scope corrected in 2.5.146: the My Account button no longer zeroes or deletes anything server-side — it only queues the command, resetting the device's own local counters and keeping the account's stored detection history; the admin tile keeps the full behavior described here.)* This had to work this way around: the PC is the lifetime counter's source of truth (every heartbeat writes the total from the device), so clearing only the website would have lasted one heartbeat. The race is closed: while the command is unacknowledged, the heartbeat handler holds the stored lifetime at 0 instead of accepting the device's old number; once the device acknowledges — it acks only after executing the reset — new counts flow again. A repeat reset while one is pending queues nothing new, and every reset is audit-logged (install.reset_stats). The customer endpoint is strictly ownership-scoped (a cross-account attempt is a 403 and queues/deletes nothing). Not reset, by design: quarantined files stay quarantined; false-positive exclusions are separate records and survive, so a file the user marked safe is never re-flagged; ML verdict history and verdict reports (internal training data) are untouched. Suite: v25144-reset-stats-test.js (52 checks, including the full reset → stale-heartbeat → ack → fresh-count sequence). - Copy: Help FAQ gains "Can I reset my threat counters back to zero?", the admin guide documents the install-detail tile, and both confirm dialogs state the scope in plain words before anything is zeroed.
- Server side of the pair with Client 2.5.121 below; nothing else changes.
Version 2.5.181
- Fix-it actions v2 — the six held actions from the approved proposal, on the shipped v1 framework (needs client 2.5.143). Same command family, same safety model, same billing rule (a fix outcome landing never creates a fee; the only charge is the repair row chosen at queue time), same roles (owner + manager; support 403), same banner-only consent with the completion wording that points the customer to their technician. The six:
fix-startup-disable:<source>:<name> / fix-startup-enable:<source>:<name> (the source is one of ten fixed tokens naming the same sources diag-startup enumerates — Run/RunOnce in HKLM and HKCU, native and 32-bit views, plus the user and common Startup folders; logon scheduled tasks have no token and are never disable targets; the client re-derives its enumeration at run time and refuses an entry it cannot find; registry values are moved to the holding key Software\AEAntivirus\DisabledStartup\<source> under the same hive and view, preserving name, type and data, and folder items are renamed in place with a .ae-disabled suffix — enable is the exact reverse and refuses honestly when nothing is held under that name or a same-named entry is already back), fix-path-remove:<entry> (the fence, pinned by tests at every layer: the client removes an entry only when (a) it is found — expanded, case-insensitive, normalized — in the system PATH or a loaded user's PATH and (b) the expanded target is re-verified NOT to exist on disk; if it exists or is not in any PATH it refuses with that reason and removes nothing; before editing it backs up the full original value into HKLM\Software\AEAntivirus\PathBackup as a timestamped set, then removes ALL occurrences from that one PATH value, preserving its registry type; entries whose variables the PC cannot resolve are refused rather than guessed at), fix-path-restore (restores the newest unused backup set, marks it consumed; no backup is an honest failure), fix-share-remove:<name> (fences: any name ending in $, plus ADMIN$/IPC$/print$ spelled out, plus any share the Shares check marks standard, plus a name that does not exist — all refused; NetShareDel on success; the outcome states plainly that the folder and its files are untouched and that re-creating the share is manual — there is no undo), and fix-dism-restore (Dism.exe /Online /Cleanup-Image /RestoreHealth through a fixed-binary runner in diag.c built exactly like the SFC one — the binary path and the whole argument list are constants, nothing from the queued command reaches the command line; ~30 s progress heartbeats, a 60-minute cap, cancel through the existing cancel path at heartbeat boundaries; verdicts in DISM's own terms — completed, image repaired / completed, no corruption found / failed, repair files not found (source needed) / failed with the error — and DISM reporting the operation was cancelled is treated as a cancellation, never a partial claim). Validation mirrors v1's three-way discipline (server remoteCommandIsValid, client cmd_parse_one, client fix.c) on one pinned matrix: bad source tokens, oversized names/entries, and ; or " in a PATH entry are rejected on every end. Panel: Disable per startup entry on the Startup card (with Enable on entries the latest fix outcomes show as currently held — the panel derives held state from the stored disable/enable outcomes, newest per entry wins, since the panel keeps no state of its own; the disable outcome card carries the same Enable button), Remove per broken entry on the PATH card plus a Restore PATH backup button while the latest path outcomes report an unrestored backup, Remove per non-standard share on the Shares card (standard rows get no button), and a DISM repair (RestoreHealth) button beside the SFC repair button; every dialog carries the v1 elements (action + target, the undo truth stated plainly — share removal and PATH restore among them — and the No charge / Create repair charge choice). The Findings page gains no new Fix mappings (its only Fix mapping remains v1's cleanup). Copy: ToS §4t and the Privacy Remote fixes bullet extended (startup holding key, on-device PATH backup + restore, share folder-untouched + manual re-create, DISM); /tech-services EN + ES fixes lists extended; help FAQ (EN + ES) extended; the admin guide gains the v2 section including the Enable-surfacing mechanism; the homepage rotates (Server 2.5.181 + Client 2.5.143 + Server 2.5.180 visible; Client 2.5.142 heads the fold). Tests: new server/tests/server/v25181-fixit2-test.js (v2 validator matrix, four-place wiring pins, fake-device round trips per command incl. the PATH fence's three outcomes, the startup disable→enable round trip, the share standard refusals C$/IPC$/$-suffixed, path backup→restore, DISM outcome mapping from canned outcome payloads, repair-attach + no-fee asserted once for a v2 action, copy pins, source scans, support 403s, inline-script regression) and extensions to the client fix/cmd tests (validator matrix in C, canned DISM verdicts pinned through the pure fix_dism_verdict mapper the diag.c runner uses).
Version 2.5.143
- The client half of Fix-it v2: six more fenced fixes in
src/fix.c. Startup disable/enable re-derive the same enumeration diag-startup reports (Run/RunOnce across HKLM + HKCU in both views, both Startup folders) and act only on entries found in it at run time: registry values move to the Software\AEAntivirus\DisabledStartup\<source> holding key under the same hive/view with name, type and data preserved, folder items are renamed with a .ae-disabled suffix, and enable restores without ever overwriting a same-named entry that is already back. PATH remove applies the fence (found-in-a-PATH + target-verified-missing, per-user expansion from each loaded hive's own profile exactly as the PATH check does, never force-loading a hive), backs the full original value up into the PathBackup key before editing, removes all occurrences, and verifies against a fresh read; PATH restore writes the newest unused backup set back byte-for-byte and marks it consumed. Share remove enumerates via NetShareEnum (dynamically loaded, the diag-shares pattern), refuses the standard set three ways (the $ suffix, the spelled-out ADMIN$/IPC$/print$, and the IPC type the Shares check marks standard), deletes with NetShareDel, and verifies the share is gone. DISM runs through the new diag_dism_restore_run bridge in diag.c (fixed %SystemRoot%\System32\Dism.exe, fixed argument list, 60-minute cap, heartbeat progress and cancel exactly like the SFC repair bridge), with the verdict mapping in a pure, host-tested function (fix_dism_verdict) so canned DISM outputs are pinned by the client test. fix.c still contains no process creation of its own (source-scanned); the family's only subprocesses are sfc.exe and Dism.exe through diag.c's fixed-path machinery. What's New names the six actions, the holding-key rule, the PATH fence and backup, the share no-undo truth, and the DISM cap.
- One button style everywhere (Austin's requested design standard). Every button on the site and in the admin panel now wears the exact look of the My Account "Download my data (JSON)" button: cyan gradient fill (
linear-gradient(92deg,#22d3ee,#0ea5e9)), dark #04121a text, 12px rounded corners, 14px/32px padding, bold 16px type, and the soft cyan glow (0 8px 26px rgba(34,211,238,.22)) that deepens on hover. The unification lives at the shared-CSS level — the old muted .btn.ghost treatment is now an alias of the standard button, so secondary and dialog Cancel/Back buttons match too — and the page-local button styles that fought the standard were normalized to the same values (the Password Vault page's blue buttons, the link-checker widget's submit button, the remote-session viewer's toolbar and file buttons, the admin pricing-catalog size override, the admin phone-width shrink). Keyboard users keep a visible focus ring (.btn:focus-visible), and disabled buttons dim with a not-allowed cursor. The two approved exceptions: destructive actions keep the identical shape in a red fill (the new .btn.danger — Delete license, Remove device, Disable 2FA, key/passkey/remote-access Revokes, and the admin Delete/Revoke/Block actions, Merge accounts included), and buttons inside dense admin tables take the same colors/radius/font at the compact .btn.sm padding so rows don't break. Reseller brand color is preserved: branded pages flow their accent through --grad, so a reseller's buttons wear the reseller's color in the standard shape — shape unified, brand color untouched. A new regression suite (v25143-buttons-test.js) pins the canonical rule values on the served pages. - Footer links lose their underlines. Footer links no longer render underlined at rest — including in Comfort view, whose blanket
text-decoration:underline for links previously overrode the footer rule (a comfort-specific carve-out now wins it back) — and hovering brightens the link text instead; the hover rule's color variable (--cyan) is now actually defined in the public :root, where several components already referenced it. Body-content links outside the footer keep their underlines. - Server-only release; the Windows client stays 2.5.120. The visual verdict is Austin's on the live site, as always.
Version 2.5.180
- Fix-it actions — the Tech tools go from read-only to nine fenced, reversible fixes (needs client 2.5.142). The 49 checks stay read-only; a second command family,
fix-*, lets an owner or manager queue a fix from the Tech tools panel for a device to apply at its next check-in: fix-service-start:<service> (starts a named service — re-validated against the device's own service list; unknown is an honest failure, a disabled service reports "service is disabled", already running is a success "already running"; waits up to 15 s and reports the verified after-state), fix-print-clear:<printer> (exact-match printer from the device's own list; winspool purge first with a queue-empty verification, fallback restarts the spooler and clears the shared spool folder and says it cleared ALL printers' queues), fix-dns-flush (DnsFlushResolverCache), fix-time-sync (W32TimeSyncNow, fallback restarts W32Time; the result carries the device epoch and the server annotates clock_skew_seconds at ingest exactly like diag-time, so "clock now within N seconds" is measured by the server), fix-cleanup:<targets> (comma-joined subset of the FIXED ids {temp, wucache, recycle} only — the client maps ids to the same paths diag-cleanup measures from its own fixed table, deletes contents, skips and counts locked files, re-measures, and reports freed MB before−after, as a floor with a partial flag when a measurement capped), fix-cleanup-winold (Windows.old as its OWN command, never a fix-cleanup target — pinned on both validators — because deleting it removes the rollback of the last Windows upgrade, which the outcome states), fix-wu-reset (stops wuauserv + BITS, renames SoftwareDistribution\Download to Download.aebak-<yyyymmddhhmmss> with a counter suffix on collision, deletes aebak backups older than 7 days, restarts both services; the outcome names the backup and its 7-day undo deadline), fix-wu-reset-undo (restores the newest backup the same way; nothing to restore is the honest failure "no backup available — it may have passed the 7-day window"), and fix-sfc-repair (sfc /scannow through the same machinery as diag-sfc minus /verifyonly, verdict words mirrored). Safety model: there is no freehand mode — command arguments are validated exactly like the diag-drive/diag-cancel precedent on both ends (service-name charset [A-Za-z0-9_.-] ≤ 64, printer name ≤ 128 with no path characters, cleanup ids from the fixed set), a source-scan test pins that the client's fix code contains no process creation at all (no CreateProcess/ShellExecute/system/WinExec — the one subprocess, sfc.exe, runs through the diag machinery) and that fix target paths come only from the fixed table, and every target is re-validated on the device against its own enumerations. The wiring is the usual discipline (server command whitelist, server remote-command validator, client parser + CMD_TYPE_MAX, both engine dispatchers, diag-cancel accepts the nine fix names), journaled on the device as fix_run BEFORE acting, cancel-aware at every boundary (a cancel posts {cancelled:true, stage_reached}, never a partial outcome), with stage progress through the existing diag-progress endpoint and results through a NEW device-auth POST /api/fix-result mirroring diag-result (deep-sanitized, >32 KB refused, fix_results table, newest 5 per action per device). Consent is the existing notice, extended: one in-progress system now covers checks and fixes (shared run registry) — a balloon when a fix starts ("a technician is applying this fix on your PC"), a persistent banner while any fix is active with the wording that a technician is APPLYING A FIX (fix wording wins over check wording while a fix is active), and ONE completion balloon when the last active run of either kind ends ("the fix on your PC has finished — please check with your technician for the next step"); the phrase "safe to restart" remains banned (source-scanned and asserted absent from both built EXEs). Billing rides the existing repairs ledger, never auto: every Fix button opens a confirm dialog naming the action, the target, the undo truth, and a billing choice — No charge (default) or Create repair charge (labor/parts/description through the existing record-repair path: POST /api/admin/commands accepts an optional repair attach, creates exactly one service_fees repair row at queue time, stamps its id onto the queued command entry so the landed result and the panel show "Repair #N attached"; the description defaults to "Fix: <friendly action> — <target>"). A fix result landing NEVER creates a fee by itself (pinned, the opposite of the diag auto-fee), and cancelling a queued fix leaves an attached repair row in place (the guide tells the admin to waive or delete it). Panel: Fix buttons on the services rows (Start, stopped/disabled state permitting), printer rows (Clear queue), the time card (Resync now), the Windows Update card (Reset update components, plus Undo when the latest wu-reset outcome names a backup), the SFC card (Run repair scan), the network card (Flush DNS), and the cleanup card (Run cleanup with the three target checkboxes pre-ticked from the latest cleanup result + a SEPARATE danger-styled Windows.old button whose confirm carries the extra rollback warning); the cleanup finding on Needs attention opens the same dialog. Fix pending/progress render like diag cards (same stage machinery + Copy customer update), outcomes render verified / ran-but-unverified / failed, and fix history lists in the device panel from fix_results. Schedules do NOT gain fixes — diagnostics only, stated in the guide (a fix is a decision a technician makes in the moment, not a recurrence). Roles: owner + manager only; support gets 403 on the queue, fix-result reads, and cancel paths (pinned). Copy: ToS gains §4t "Remote fixes" and Privacy a Remote fixes bullet naming every action, the banner, the fixed-target design in plain words ("a fix command can only ever choose from a short list of known actions — it cannot run programs or open files on the device"), the Windows.old rollback consequence, and the 7-day WU backup; /tech-services (EN + ES) gains the remote-fixes list with technician-initiated framing; the help FAQ (EN + ES) answers "Can a technician fix something on my PC remotely?"; the admin guide gains the full fix section; the homepage rotates (Server 2.5.180 + Client 2.5.142 + Server 2.5.179 visible; Server 2.5.178 heads the fold) and gains a Remote fixes card in Technician services. Tests: new server/tests/server/v25180-fixit-test.js (validator matrix incl. winold-inside-fix-cleanup rejected on both ends and path-like args rejected, four-place wiring pins, per-action fake-device round trips incl. the already-running service, the spooler fallback, Windows.old not-present, the time-sync server skew annotation, the wu-reset → undo round trip, repair-attach = exactly one row with linkage, no-charge = none, result-lands-alone creates no fee, queued-fix cancel leaves the repair row, support 403s, notice copy pins, the fix.c source scans, inline-script regression on every page) and client tests server/tests/client/fix-test.c/.sh + new cmd-test cases; client 2.5.142 adds src/fix.c (the nine actions, the fixed target table, the shared-notice integration).
Version 2.5.142
- The client half of Fix-it actions:
src/fix.c, the nine fenced fixes. Every fix journals fix_run before acting, registers in the shared run registry (so the tray notice covers it and diag-cancel can stop it), posts stage progress through the existing progress endpoint, and posts its outcome to the new /api/fix-result. The action set: service start (own-enumeration validation, disabled/unknown honest failures, 15-second verified wait), print queue clear (winspool purge + verify, spooler-restart fallback that names the all-printers consequence), DNS flush, clock resync (W32TimeSyncNow with a service-restart fallback, epoch post for the server skew annotation), itemized cleanup (the fixed target table only — temp / update cache / recycle bins, emptied via SHEmptyRecycleBin; before/after re-measurement, locked files skipped and counted), Windows.old removal (its own command, freed space + the rollback consequence stated), Windows Update reset + undo (the Download.aebak-<timestamp> rename-aside, 7-day purge, both services restarted, states reported), and the system file repair scan (through the existing SFC machinery, without /verifyonly). The notice integration treats an active fix exactly like an active check for the banner, with fix wording while any fix is active, and still emits exactly one completion balloon when the last run of either kind ends. What's New in the client names the nine fixes, the short-known-actions safety model, and the Windows.old rollback warning.
- Account merge tool (admin, owner-only). Customers → Accounts gains Merge accounts:
GET /api/admin/accounts/merge-preview?from=&to= returns the exact counts that would move — enumerated from the actual schema: licenses (with the installs counted through their keys), paid add-ons, support messages (that table keys on email, so messages follow the address), customer API keys, family links (owned + memberships + invites), referrals — plus warnings computed from the data: the source's own Stripe customer ("subscriptions stay under the old Stripe customer — handle in Stripe; local records move"), add-on conflicts (destination's row kept), enabled 2FA and passkeys (credentials never transfer), family links and referrals *between* the two accounts (dropped — an account cannot link to or refer itself), and duplicate referral codes (destination's kept). One schema finding is stated in the preview rather than hidden: verdict_reports has no account_id and no email (anonymous tool submissions), so no verdict report can be attributed to either account and none moves. POST /api/admin/accounts/merge requires typing the source email, then moves everything in a single transaction, revokes the source's sessions, deletes its 2FA/passkeys/tokens, and closes it as status='merged' with a merged_into pointer; its login is refused (after the password verifies, so the message can't probe for emails) with "This account was merged into another account." Audit entries are written for both sides (account.merge.out / account.merge.in) and a merge note lands on both accounts. Guards: self-merge, already-merged source, and merging into a merged/deleted account are all refused; the routes are behind requireOwner AND in the helper-admin owner-only denylist next to refunds and backups. - Site maintenance mode. System → Maintenance gains Site maintenance mode (owner-only toggle + custom message, audited as
site.maintenance, effective instantly — the gate reads the setting per request). While ON, the public pages a visitor browses (/, /subscribe, /help, /download, /business, …) answer 503 with a branded maintenance page carrying the admin's message. Always exempt, by design and by test: all /api/* (heartbeats, registration, detections, checkers, the Stripe webhook, admin APIs), /admin, the reseller portal, /status, /account + /login (existing customers unaffected), /updates/*, and the installer file routes (/download/file/*, /download/*.exe), plus PWA plumbing and brand icons. The gate fails open — a settings error can never take the service down. (Not to be confused with the older quiet-hours "maintenance windows", which only suppress email and are unchanged.) - Off-site backup copies. A nightly
offsite-backup job (after 03:00; listed in the jobs viewer as Run-Now-safe — it only reads the backup and uploads it) copies the latest automatic backup to an S3-compatible bucket configured purely by env: OFFSITE_S3_BUCKET / OFFSITE_S3_KEY / OFFSITE_S3_SECRET, optional OFFSITE_S3_REGION (default us-east-1) and OFFSITE_S3_ENDPOINT (for B2/R2/MinIO-style services). The single PutObject is signed by hand with AWS SigV4 (HMAC-SHA256 via node:crypto, path-style addressing — the form every S3-compatible service accepts): no new npm dependency, so nothing had to be mirrored into the root package.json. Every attempt is recorded in offsite_backup_runs, and Server Health shows "Off-site copy: <date> — OK / Failed" next to the restore-verification line — or "Not configured" until the env vars exist, in which case the job skips cleanly. - Server-only release; the Windows client stays 2.5.119. Nothing for Austin to configure unless he wants off-site backups (env vars above + a redeploy).
Version 2.5.179
- Tech tools workflow round (server only — the client stays 2.5.141 and no manifest changes). The 49-check catalog is finished; this release builds the workflow around it. Findings alerts: when a diagnostic result lands and the existing findings engine flags the device, the serious findings notify the shop immediately through the existing notification machinery (admin notification center + Web Push + the normal mail path — sendMail journals every attempt into
mail_log, so with SMTP unconfigured the attempt is recorded there, exactly like alert-rule mail). The serious set is exactly six of the engine's rules: drive health bad (diag-drive drives[].level), SSD wear ≥ 80% or level bad (diag-ssd), bugchecks present (diag-crashes), AV problems — more than one active scanner or signatures out of date (diag-avstate), Windows not activated (diag-license), and pending reboot (diag-reboot). The informational findings (cleanup ≥ 20 GB, Windows 11 minimums unmet, a non-standard share, clock drift) list on Needs attention but never alert. Dedupe state lives in diag_finding_alerts (install + tool + a fingerprint naming the cause set, from the same stored fields the rules read): re-posting the same finding does not re-alert; the alert re-arms only after a later real result for that tool shows the finding gone — an error or cancelled result neither fires nor clears, since it says nothing about the finding. Each fired alert is audit-logged (diag.finding_alert) with the device label, tool label, and one-line summary. Create repair from a finding: Needs attention rows (and the findings rows on the printable device report, via a deep link) gain a Create repair action that opens the existing Record-a-repair form for that device — same endpoint, same service_fees ledger, no second repair path — with the device preselected, the description prefilled with the finding's server summary line (editable before save, per the form's rules), and the finding's own result preselected under "Based on diagnostic" (the findings API now carries result_id + summary for exactly this); saving lands the admin on Customers → Service fees where the new repair shows. Customer results view: My Account → Billing gains Checks by your technician (GET /api/account/diag-results, own account only — installs follow their license_key into licenses.account_id, the same join installAccountId uses): per device, the newest 20 results with the localized tool name (EN/ES label tables, following the page language), the date, and a customer-safe one-line summary — safe by construction, since the only strings are the fixed "couldn't run on this device" / "cancelled" phrases and the server's shared summary line (the same text Copy-customer-update produces); no payload fields, no findings or severity language cross over. Health trends: a new diag_history table keeps, for every landed result, the server summary line plus a few numeric metrics extracted only from fields the renderers already read — diag-ssd highest wear_pct, diag-battery health % (100 − wear_pct) + cycle_count when reported, diag-time clock_skew_seconds (the ingest annotation), diag-cleanup total_reclaimable_mb, diag-bootperf newest boot_ms. diag-drive (health is words, not a number) and diag-space (its payload carries folder sizes only — no volume free-space fields exist) store the summary only; nothing is invented for them. Rows are capped at the newest 500 per install+tool (rolling delete at insert). Each panel card gains a History toggle (new GET /api/admin/installs/:id/diag-history) with the timeline and, for metric tools with ≥ 2 points, a self-contained inline-SVG sparkline; the printable device report gains a Trends block only where a metric moved materially (SSD wear up ≥ 5 points, or battery health down ≥ 5, across the kept window), phrased as plain facts ("SSD wear rose from 71% to 83% over 6 checks"). Retention: diag_history is deliberately registered in neither cleanup framework (not in PRUNABLE_STORES, not in RETENTION_CLASS_SPEC) — like FP exclusions and licenses it is never auto-purged, and the admin guide + Data cleanup copy say so; the 500-per-check cap is its bound. Copy: the admin guide gains the workflow paragraph (alert set + dedupe, create-repair, customer view, trends + retention); Privacy and ToS §4s state plainly that account holders can view check summaries in My Account and that serious findings alert the shop; /tech-services (EN + ES) names the new customer section and the automatic alerts; the homepage rotates (Server 2.5.179 + Server 2.5.178 + Client 2.5.141 visible; Server 2.5.177 heads the fold). Tests: new server/tests/server/v25179-workflow-test.js; the soak baseline is regenerated for 2.5.179 (admin GET endpoints +1: diag-history).
Version 2.5.178
- Tech tools round 8 — the final round: three more read-only checks, and the catalog is complete at 49 checks (counted from the server's own tool table; the admin guide, this README, and the /tech-services page all state the same total, and the release suite asserts the stated number equals the table). Same four-place wiring (server command whitelist, server remote-command validator, client parser + cancel names, both client dispatchers via
diag_is_tool), bare-form commands only, journaled on the device before collecting, with progress posts, boundary cancel checks, the on-device in-progress notice, guide-duration estimates, and the standard fee rules (a completed run can bill the configured diagnostic fee; an error or a cancellation never does; a clean-empty informative result is a completed run and bills). The three: Windows 11 readiness (diag-win11) — the device reports FACTS ONLY (Windows edition + build from the CurrentVersion key, the TPM specification token from Win32_Tpm — the raw SpecVersion string's leading token, omitted when no TPM answers — Secure Boot from the same UEFISecureBootEnabled value the posture check reads, with an absent key reading "not reported" for legacy BIOS, processor name/cores/MaxClockSpeed from Win32_Processor, memory from GlobalMemoryStatusEx, and the system drive's total/free space); the verdict is scored by the server renderer against Microsoft's published minimums (TPM 2.0 present, Secure Boot on, RAM \≥ 4 GB, system drive \≥ 64 GB, CPU \≥ 2 cores and \≥ 1000 MHz), per item Meets / Does not meet / Not reported, with the overall line "Meets every minimum we can check" only when every checkable item passes. The CPU line never claims Microsoft-list support: when speed and cores pass it reads "meets the speed and core minimums — Microsoft's supported-CPU list is the final word". DirectX/WDDM is deliberately NOT scored (not reliably checkable) — there is no graphics line at all. Cleanup candidates (diag-cleanup) — bounded walks (the diag-space walker: one shared 250,000-entry / 120-second / cancel budget across all targets, partial flag) of Windows.old, the Windows Update download cache (SoftwareDistribution\Download), the Recycle Bin across volumes ($Recycle.Bin per volume root, summed into one item), Windows\Temp, and the system temp folder (skipped when it resolves to the same folder as Windows\Temp, so nothing is double-counted); payload {items: [{name, size_mb, reclaimable}], total_reclaimable_mb, partial} with missing targets simply absent (no Windows.old is normal). The component store (WinSxS) is reported separately with reclaimable: false and the honest label "size on disk — not a reclaimable amount; the component store shares files with Windows"; the renderer leads with the reclaimable total. The check measures only — it deletes nothing (pinned by a source scan of the collector). Bluetooth devices (diag-bluetooth) — paired devices from the Bluetooth port driver's own records (HKLM\SYSTEM\CurrentControlSet\Services\BTHPORT\Parameters\Devices: one subkey per paired device, named by its address, the device name in its Name value), with a coarse kind decoded from the class of device (phone / computer / audio / peripheral / other / unknown) merged in from the classic enumeration API by address (dynamic load of bthprops.cpl, the tray.c pattern; paired inventory only — fIssueInquiry is FALSE, no inquiry is ever issued), the address formatted as a MAC and collected/disclosed exactly like diag-adapters' MACs, capped at 50, shipped as a paged stream (device_pages, pages of \≤40, with the true total) because the result sanitizer slices arrays at 40. Battery: a battery_pct is included ONLY from the PnP device property Windows itself caches for Bluetooth devices ({104EA319-6EE2-4701-BD47-8DDBF425BBE5} pid 2 — the value Settings shows), matched by extracting the address from the PnP device instance ID (a pure, host-tested helper), included only on a unique-address match with a 0–100 value, and dropped when two cached values for one address conflict — a battery level is never estimated. No radio (neither the enumeration API nor the BTHPORT key answers) is a clean informational result ({radio: false, total: 0} — "no Bluetooth radio found"), as is a radio with nothing paired; both bill as completed runs. - Battery extension:
diag-battery also reports cycle_count from root\wmi BatteryCycleCount (CycleCount) when the class and value exist; when they don't, the field is omitted and everything else is byte-identical to before. The panel adds a "Charge cycles: N" line only when the field is present, and the one-line summary appends the count only when present — previously stored results render byte-identically, pinned by keeping the old-shape fixture as a regression next to the extended one. - Findings page gains two rules (same stored-fields-only discipline, source field cited in code): a readiness result whose server-side score (
diagWin11Unmet over tpm_spec / secure_boot / ram_gb / system_drive_total_gb / cpu_cores / cpu_mhz) fails any minimum produces an informational finding naming the unmet items — an absent or not-reported field (other than the TPM, whose absence is itself the failure) is an unknown, never a failure; and diag-cleanup total_reclaimable_mb \≥ 20480 (20 GB) produces a finding reading "about N GB can be safely reclaimed". The served page's diagWin11Score and the server's diagWin11Unmet are the same rules in two places by necessity (the panel renders client-side, the findings/report compute server-side); the diagrender suite pins both against one canonical fixture. - Copy: Privacy and ToS \§4s name the round-8 fields in plain words (the readiness facts, cleanup folder sizes — folder sizes only, never file names, Bluetooth device names + addresses, battery charge cycles); /tech-services (EN + ES) lists the three checks and states the catalog is complete at 49 checks; the homepage rotates (Server 2.5.178 + Client 2.5.141 + Server 2.5.177 visible; Client 2.5.140 heads the fold) and its Technician services card covers all 49 checks; the admin guide gains the round-8 paragraph stating the catalog is complete. Schedules and the Full checkup are unchanged — the new tools join schedules automatically through the same tool table (verified by test), and the checkup composition stays the bench-intake eight. Tests:
v25174-diagrender-test.js extended with the three canonical round-8 fixtures plus the battery cycle-count pair (49 tools), new server/tests/server/v25178-diag8-test.js (four-place guards, validator matrix incl. cancel targets and :arg rejection and support 403s, fake-device round trips for all three tools + the battery extension, the win11 scoring matrix incl. the CPU wording and the no-graphics-line pin, cleanup totals + the WinSxS label, bluetooth paging + battery-cell honesty, findings rules firing and not firing, fee pins — no-radio bluetooth bills, error payloads don't — schedules accepting a new tool, the stale-id/catalog-count copy assertions), new client tests server/tests/client/diag9-test.c/.sh (the TPM spec token, Bluetooth kind/address/name helpers incl. instance-ID extraction, the cleanup-deletes-nothing source pin, wiring guards, both-arch MinGW syntax, version floor, What's New entry, and the 'safe to restart' absence). The soak baseline is regenerated for 2.5.178.
Version 2.5.141
- The client half of Tech tools round 8 — the final three checks, completing the catalog at 49. Three new collectors in the diag worker (dispatched through the same shared path as the previous 46, so the tray balloon + persistent banner notice covers them automatically): win11 (registry + TPM-namespace + processor facts only — no verdict is computed on the device), cleanup (the diag-space bounded walker over the leftover locations, sizes only), bluetooth (the BTHPORT pairing records + the enumeration API's classes of device + the cached PnP battery property, all read-only; the enumeration API is loaded dynamically exactly like the tray's own Bluetooth code, and the battery property reader uses the SetupAPI surface the codebase already links). The battery collector also reads the charge-cycle count from
root\wmi BatteryCycleCount when the PC reports it, and omits the field silently when it doesn't. The pure helpers of the round (the TPM spec token, the class-of-device decode, the address validation/formatting, the PnP instance-ID address extraction, the registry Name cleanup) live in a new unit-tested block (DIAG8-TESTABLE) so the tricky logic is host-tested, not just compiled. What's New in the client names the three checks, the facts-only readiness rule, and the Bluetooth battery honesty note, and notes the set is complete at forty-nine checks.
- Instant threat alerts (opt-in). A new account preference
threat_alerts (default OFF, toggle in My Account -> Security) emails the account owner the moment one of their devices reports a confirmed threat: device label/nickname, threat name, file path, severity, and the action taken (the fleet on-detection action - quarantined, or deleted when the policy says so). "Confirmed" is enforced server-side in threatAlertMaybe: heuristic reports (an AI.Heuristic* name and/or a positive ai_score) are ML suspicion and never alert, and a detection whose path is already on the device's false-positive exclusion list never alerts. Storm cap: at most one alert email per account per hour (threat_alert_state); further confirmed threats inside the hour only increment a counter, and the next alert appends "Plus N more confirmed threats stopped on your devices since the last alert." Sends go through sendMail (kind threat-alert), so every attempt lands in the email log. - Family weekly digest. A new
family-digest job (checked every 15 minutes, fires Mondays, listed in the jobs viewer as NOT Run-Now-safe because it sends real email; owner trigger POST /api/admin/family-digest/run) emails family owners - accounts that own family_members links and have not opted out (toggle beside the family sharing controls on /account; preference defaults ON) - a per-device summary of the last 7 days: threats blocked (exact detections counts inside the window), last seen, last scan times, and a warning line for any protection module (real-time / firewall / web / advanced) the latest heartbeat reports off. Data-availability finding: the server records per-install detection rows, heartbeats, module states and last_quick/last_full timestamps, but per-install scan *counts* for a window are NOT recorded (installs.scanned is a lifetime counter only) - so the digest reports last scan times and never a weekly scan count; nothing is estimated. Send rule: a digest goes out only when there was at least one detection in the window or a protection-off warning; a fully quiet week sends nothing. - Sign-in history on My Account. The Security tab gains a Sign-in history card fed by
GET /api/account/login-history: the account's own rows from the 2.5.124 login_attempts journal (customer surface, matched on the account email - admin/reseller rows and other customers' rows are never returned), newest first, capped at 20. The "This wasn't me" action (POST /api/account/login-history/not-me, audited) deletes every session on the account - including the current one - and routes the browser to /login?revoked=1. There is no self-service password-change or force-change flag in the product today, so a full revoke plus a fresh sign-in (and a support password reset if needed) is the response; the security tab copy says exactly that. - Server-only release; the Windows client stays 2.5.119. Nothing for Austin to configure (SMTP still required for any email, as before).
Version 2.5.177
- Tech tools round 7 — eight more read-only checks (46 total), all wired in the usual four places (server command whitelist, server remote-command validator, client parser + cancel names, both client dispatchers via
diag_is_tool), bare-form commands only, journaled on the device before collecting, with progress posts, boundary cancel checks, the on-device in-progress notice, guide-duration estimates, and the standard fee rules (a completed run can bill the configured diagnostic fee; an error or a cancellation never does; a clean-empty informative result — no saved Wi-Fi networks, no backup activity in 30 days — is a completed run and bills). The eight: Remote access inventory (diag-remote) — Remote Desktop from the Terminal Server registry keys (the fDenyTSConnections switch, the RDP-Tcp listener port, the NLA flag) plus the TermService state from the service control manager, the Remote Assistance invitation policy (fAllowToGetHelp / fAllowFullControl), and the known remote-access tools matched by ONE client-side name table (TeamViewer, AnyDesk, RustDesk, LogMeIn, GoToMyPC, GoTo, Splashtop, Chrome Remote Desktop, RealVNC, TightVNC, UltraVNC, ConnectWise/ScreenConnect, DWAgent, Zoho Assist, BeyondTrust, Parsec) across the Uninstall registry views, running processes, and services — entries {name, how: installed|running}, running wins when both, capped at 30; the panel leads with the RDP state and any tool running right now, and presence is reported, never judged. Firewall rules audit (diag-fwrules) — enabled inbound ALLOW rules from the Windows Firewall COM API (INetFwPolicy2, linked like the WMI COM precedent — no dynamic-load precedent exists for COM in this codebase), entries {name, protocol (decoded from the NET_FW_IP_PROTOCOL number), local_ports, remote_scope: any|specific, profile (decoded)} emitted as the paged stream, capped at 200 with the true total; the panel and the summary flag rules exposing ports 3389/445/5900/22/23 to any address, computed from the stored fields; COM/API failure is an honest error payload (no fee). Power & sleep settings (diag-power) — powrprof loaded dynamically per the codebase's dynamic-load precedent (PowerGetActiveScheme / friendly name / AC+DC value indexes for sleep-after, display-off, hibernate-after — seconds converted to whole minutes, 0 = never — and USB selective suspend on AC), fast startup from HiberbootEnabled, hibernate availability from HibernateEnabled; a flat payload; a plan or value that cannot be read is omitted, never guessed. Browser settings check (diag-browserset) — user profile directories enumerated ON DISK under C:\Users (the service runs as SYSTEM; file access, not registry): Chrome + Edge Preferences per profile (the flat scanner approach — one key at a time, never a whole-document parse, because the result sanitizer stores objects past depth 4 as empty) for the homepage, the homepage_is_newtabpage flag, the startup-page count and first three, and the default search provider's name plus the HOST of its template URL only (the template itself, with its query parameters, never leaves the PC); Firefox via profiles.ini → prefs.js for the homepage pref, with its search reported as not read (stored compressed) because Firefox keeps it in search.json.mozlz4; caps 8 users / 8 profiles; zero browsers found is a clean informational result; the panel lists values plainly and makes no safe/unsafe judgment — the only cross-profile fact flagged is that values differ. Environment & PATH check (diag-envpath) — the system PATH expanded and every entry existence-tested, per-user PATHs from loaded HKEY_USERS hives only (never force-loaded; profiles not loaded counted as skipped via ProfileList, the hosts precedent), duplicates counted case-insensitively with the comparison done on the unexpanded-position-independent normalized form, the system TEMP/TMP targets tested, and the presence of SystemRoot/windir/ComSpec reported; per-user entries still holding %vars% after expansion are duplicate-checked but not existence-checked, because their variables resolve in the user's session, not the service's. Saved Wi-Fi networks (diag-wifiprofiles) — per interface, WlanGetProfileList + the profile XML (SSID/auth/cipher/auto-connect parsed by unit-tested pure helpers; the wlanapi library the adapters check already links), weakest security first (open/shared auth or WEP/none cipher), capped at 100, paged; no wireless or no profiles is a clean informational zero; no passwords or keys are ever read. Displays & monitors (diag-displays) — active displays from EnumDisplayDevices, each monitor's EDID from its Enum\DISPLAY\<instance>\Device Parameters registry entry parsed by a unit-tested pure EDID reader (manufacturer from the packed ID, model/serial from the 0xFC/0xFF string descriptors with the product code as the honest fallback, manufacture year, physical size → diagonal inches at one decimal, native resolution from the first detailed timing), current resolution + refresh from EnumDisplaySettings matched by enumeration order; capped at 8; serials collected under the USB-history precedent and disclosed in Privacy/ToS. Backup status (diag-backup) — the existing dynamically loaded wevtapi machinery over the last 30 days: Microsoft-Windows-Backup starts/successes (the channel's documented lifecycle event IDs) and failures (the channel's error-level records), the File History engine channel the same way, VSS + SWPrV service states, and per-volume shadow storage (Win32_ShadowStorage with the volume letter resolved through Win32_Volume, omitted when it cannot be resolved); zero activity in 30 days renders "no backup activity recorded in the last 30 days" and bills; a Backup channel that cannot be queried at all is an error payload (no fee). - Fleet findings — "Needs attention" (server only, from stored results). The Tech tools tab opens with a findings card computed from the latest stored result per install+tool. Every rule reads only fields the current renderers already read, with the source field cited in a code comment: drive health level bad (
diag-drive drives[].level), SSD wear ≥ 80% or level bad (diag-ssd disks[].wear_pct / level), pending reboot (diag-reboot pending === true), more than one active AV or signatures out of date (diag-avstate products[].scanner / signatures), a non-standard share present (diag-shares flattened share_pages, standard === false), bugchecks present (diag-crashes bugcheck_count > 0), clock drift over 120 s (diag-time clock_skew_seconds, the server-ingest annotation), and Windows not activated (diag-license products[].status !== 'Licensed'). Each finding names the device (a link that preselects it in the panel), the check, the panel's own one-line summary (rendered client-side by the same diagSummary), and the result's age; devices with no results are counted separately ("No checks run yet: N"). There is deliberately no "Windows updates pending" rule: diag-winupdate reports its pending count as the string not reported by design (round 2's never-estimate rule), so no real field exists to derive one from — noted here and in the code rather than invented. Owner + manager only, like the whole category (the route is not on the support allowlist). - Printable device report.
GET /admin/device-report?install=<id> (admin session + explicit owner/manager gate) renders a print-friendly page: the shop header, device identity, a spec block from the latest diag-spec result when one exists, this device's findings (the same rule engine), and the latest one-line summary + date for every check that has a result, with @media print styling (white, no navigation/buttons). Opened from buttons on the device's Tech tools panel and on the install page's compact card. Honest plumbing note: the codebase has no single business-info source — the header uses the same values the public About page and the /tech-services page print (Austin's Electronics, 1381 Tomahawk Hwy Suite B, Harrells, NC 28444, 910-214-3670) and the code comment says the three places change together; and because the panel's diagSummary lives in the served page's script, the report carries a compact server-side summary function mirroring it field-for-field (guarded by the new suite for the round-7 tools). - Scheduled recurring checks. New
diag_schedules table (install, tool — any of the 46 commands or the literal checkup, cadence weekly|monthly, next/last run, enabled, created-by) with list/create/toggle/delete endpoints under /api/admin (owner + manager; support 403 by the default gate). The Tech tools panel for a selected device lists its schedules and adds new ones (tool dropdown including Full checkup, cadence), with copy stating plainly that scheduled runs bill like manual runs when results land and run on the device's next check-in after the due time. Firing rides the codebase's existing in-process job mechanism (jobRun + JOB_DEFS + a 15-minute setInterval, overlap-guarded): due schedules queue through the same queueCommand path (checkup = the same 8-command set, mirrored server-side with a comment), next_run advances from the later of the scheduled time and now (weekly +7 days; monthly +1 calendar month with the day clamped to the target month), last_run_at is set, and create/delete/fire are audit-logged; one schedule's failure never blocks the others. The job is listed on the Scheduled jobs card with run: false — it queues billable work, so it is never run by hand from that card, the same treatment as the email jobs. - Copy: Privacy and ToS §4s name the round-7 fields in plain words (remote-tool names and running state, firewall rule names/ports/scopes, browser homepages and startup addresses plus search provider name/host, PATH entries, Wi-Fi network names, monitor serials, backup history); /tech-services (EN + ES) lists the eight checks plus scheduled checks and the printable report; the homepage rotates (Server 2.5.177 + Client 2.5.140 + Server 2.5.176 visible; Client 2.5.139 heads the fold) and the Technician services card covers all 46 checks and the three features; the admin guide gains the round-7 paragraph. Tests:
v25174-diagrender-test.js extended with the eight canonical round-7 fixtures (46 tools), new server/tests/server/v25177-diag7-test.js (four-place guards, validator matrix incl. cancel targets, :arg rejection and support 403s, fake-device round trips for all eight, the findings rules incl. the no-results count and devices with results but no findings, the report route's content + print CSS + role gates, schedules CRUD + a due-sweep fire incl. checkup = 8 commands, monthly day-clamping, toggle/delete, fee pins — backup zero-activity bills, error payloads don't — and copy pins), new client tests server/tests/client/diag8-test.c/.sh (the remote name table's match order, firewall word decoders and port coverage, power unit conversion, the Chromium JSON finders incl. the homepage/new-tab near-key trap and host-only search extraction, the prefs.js reader, PATH split/dup rules, the WLAN XML reader, the EDID reader, the backup event classifier, wiring guards, both-arch MinGW syntax, version floor, What's New entry, and the 'safe to restart' absence). The soak baseline is regenerated for 2.5.177.
Version 2.5.140
- The client half of Tech tools round 7. Eight new collectors in the diag worker (dispatched through the same shared path as the previous 38, so the tray balloon + persistent banner notice covers them automatically): remote (registry + SCM + the three-source tool census), fwrules (the firewall COM API; enabled inbound allows only, paged under the post budget with the true total), power (powrprof dynamically loaded; a value that cannot be read is omitted, never guessed), browserset (on-disk profile enumeration; flat one-key-at-a-time JSON finders — no whole-document DOM, so nothing depends on document key order and no value is ever read past its own key), envpath (system + loaded-hive user PATHs; unloaded profiles counted, never opened), wifiprofiles (profile XML per interface; only names and security types — no key material exists anywhere in the code path), displays (EDID from the registry, current settings from the display API, matched by enumeration order), backup (the events machinery with a shared 2000-record / 60-second budget across both channels; the shadow-storage half is best-effort and its failure omits only itself). The pure parsing/matching core of all eight lives in a new unit-tested block (name-table matching, protocol/profile words, port coverage, seconds→minutes, the JSON finders, the URL host extractor, the prefs.js reader, PATH splitting/dedup, the WLAN XML reader, the EDID reader, the backup classifier) so the tricky logic is host-tested, not just compiled. What's New in the client names the eight checks, the search-address honesty note, and the Wi-Fi no-passwords note.
- Chargeback & dispute watch. Stripe disputes are no longer buried in the webhook log:
charge.dispute.created / charge.dispute.closed events land in a new stripe_disputes table (dispute id UNIQUE — Stripe retries update the row, never duplicate it) and Customers → Revenue gains a Chargebacks & disputes list: amount, reason, status, evidence due date, and the customer where resolvable through the charge (best-effort — an unresolvable dispute still records, marked unresolved). The owner is alerted by email exactly once, on the first insert only, so retries can never re-alert; the alert lands in the mail log like every other send. - Backup restore verification. A backup that has never been restored is a hope, not a backup: the new weekly
backup-verify job takes the latest automatic backup, extracts it into a scratch directory (never in place; the real file is only ever read), opens the restored database read-only, runs PRAGMA integrity_check plus a core-table sanity pass (accounts / licenses / installs / detections present, row counts valid), and records the verdict in backup_verifications. Server Health now shows "Last backup restore test: <date> — Passed/Failed" with the file name and detail; a failure also writes an audit entry (backup.verify.fail). "No backup found" is recorded as a failure state rather than crashing, encrypted backups are decrypted with the configured passphrase, and the job is honestly Run-Now-safe (it writes only to scratch). - Reseller payout statements. The reseller portal gains Payout statements: per calendar month, totals computed from the existing
reseller_commissions accrual records — gross, commission, paid vs accrued — with the per-sale rows behind each total and a CSV download. Statuses are labeled honestly: a commission reads paid only when its row sits in a completed payout, otherwise accrued (earned, not yet paid). Statements are session-scoped to the signed-in reseller — isolation is enforced by the query itself, and a reseller with no commissions gets a clean empty statement. Server-only release; the Windows client stays 2.5.118. Nothing for Austin to configure.
Version 2.5.176
- Tech tools round 6 — nine more read-only checks (38 total), all wired in the usual four places (server command whitelist, server remote-command validator, client parser + cancel names, both client dispatchers), bare-form commands only, journaled on the device before collecting, with progress posts, boundary cancel checks, the on-device in-progress notice, guide-duration estimates, and the standard fee rules: a completed run can bill the configured diagnostic fee; an error, a cancellation, or a no-data outcome never does — while a clean-empty result (no crashes recorded, no restart pending) is a completed run and bills like the Problem devices zero before it. The nine: Driver inventory (
diag-drivers) — every Win32_PnPSignedDriver record (device, provider, version, DMTF date normalized to ISO, INF name, signed flag + signer), capped at 300 collected entries with the true total counted past the cap and emitted as a paged stream (the result sanitizer caps arrays at 40 — the round-5 paging pattern keeps every entry at depth ≤3); the panel flags unsigned drivers and drivers older than five years at the top and the summary counts total / unsigned / oldest year. BitLocker & TPM (diag-bitlocker) — per-volume conversion + protection words from Win32_EncryptableVolume (reusing the posture check's decoders) and the TPM state from Win32_Tpm; key material is never read — no protector enumeration, no recovery value anywhere, pinned by a source scan (the method name appears nowhere in the client diagnostic source) and a fixture asserting the stored payload carries no key/protector fields; a refused encryption namespace returns an honest error payload (no fee). Pending reboot & servicing (diag-reboot) — CBS RebootPending and Windows Update RebootRequired key presence, PendingFileRenameOperations (count + up to 20 paths), and a pending computer rename; the panel leads with the verdict and zero-pending is a first-class clean result. Antivirus products (diag-avstate) — SecurityCenter2 products with productState decoded to words using the documented bit layout (bit 0x1000 = real-time scanner on; bit 0x10 set = signatures out of date — validated against known healthy/stale sample values), plus a flat Defender object from MSFT_MpComputerStatus (service/AV/real-time on, signature version and its age in days computed on the client); the panel flags more than one active product, out-of-date signatures, and a present-but-disabled Defender. App reliability history (diag-reliability) — Win32_ReliabilityRecords for 30 days grouped by product (crashes/hangs/last seen, top 25); only Application Error / Application Hang sources count — Windows Error Reporting records duplicate the same failures and are excluded by construction, as are installer/update sources, and blue screens stay in Crash history; a clean 30 days renders as a first-class clean result. Memory check (diag-memory) — the last MemoryDiagnostics-Results event via the round-4 wevtapi path (verdict from the event ID — 1101/1201 passed, 1102/1202 failed — with the provider's own message text as the only fallback; never-run stated plainly and clean), page files from Win32_PageFileUsage, and memory pressure from GlobalMemoryStatusEx; the modules themselves stay in Hardware inventory. Shared folders (diag-shares) — NetShareEnum (netapi32 loaded dynamically; no new link dependency) at level 2 so each share carries its path and current connection count, falling back to level 1 only when access is denied (path/uses then omitted, never guessed); $-suffixed and IPC shares are marked standard, the list is sorted non-standard first, capped at 100, paged like the drivers list. Clock & time sync (diag-time) — W32Time service state (via the service control manager), the registry sync Type/NtpServer, the last successful sync from Time-Service events 35/37 with its source, and the device's own UTC clock reading (client_epoch); the skew is measured by the server at ingest against its own clock and annotated onto the stored result as clock_skew_seconds (the after-cancel-requested annotation precedent — a server-side fact, not client payload); under 120 seconds the panel reads "clock in sync", larger drift reads the amount and direction. Performance snapshot (diag-perf) — formatted processor/system/disk counters and available memory, plus top-10 processes twice from Win32_Process: by cumulative CPU seconds (kernel + user time — labeled "most CPU since boot", never as current load) and by working set. - Full checkup preset (server UI only, no new command). Tech tools gains a Run full checkup button on the single-device panel and a Full checkup option in the multi-run card: the bench-intake set queued in order — Health snapshot, System spec sheet, Drive check (all drives), Drive reliability, Disk space, Security posture, Problem devices, Crash history. System file verification is deliberately excluded (it is the 30-minute heavyweight; the admin guide and both Tech tools copy blocks say so). Multi-run outcomes render per device as a queued count out of 8 with any failures named; one failed command never blocks the rest of the set, and fees follow each check's normal landing rules. The admin guide's Tech tools section gains the round-6 paragraph; the Privacy policy and Terms §4s enumerate the new collected fields in plain words (including the "no BitLocker recovery keys, ever" statement); the /tech-services pages (EN + ES) list the nine checks and the checkup; the homepage rotates (this is a customer-facing release) and its Technician services card enumerates all 38 checks. Tests: new
server/tests/server/v25176-diag6-test.js (four-place guards, validator matrix incl. cancel targets and arg rejection, fake-device round trips for all nine, clock-skew measured at ingest and rendered both ways, bitlocker fixture with no key fields, fee rules — clean-empty bills, error/cancelled never — the checkup queued as 8 in order per device in both modes with one failure not blocking others, copy pins) and new client tests server/tests/client/diag7-test.c/.sh (productState decode vectors, DMTF dates, share words, memory verdicts, the reliability source predicate, time-source extraction, the never-read-key-material source pins); v25174-diagrender-test.js gains the nine canonical round-6 fixtures. The soak baseline is regenerated for 2.5.176.
Version 2.5.139
- The client half of Tech tools round 6. Nine new collectors in the diag worker (dispatched through the same shared path as the previous 29, so the tray balloon + persistent banner notice covers them automatically): drivers (paged emission under the 32 KB post budget with a truncated flag and the true total), bitlocker (status properties only — the volume path reuses the posture check's read-only namespace walk; TPM state from its own namespace), reboot (registry evidence only — no WMI), avstate (SecurityCenter2 + the Defender namespace; the raw
productState number is decoded on the device and never sent), reliability (30-day window computed against the device's clock; window totals counted independently of the top-25 list), memory (event-log verdict + WMI page files + GlobalMemoryStatusEx), shares (dynamic netapi32, sorted non-standard first), time (service state from the SCM, sync config from the registry, last sync from the event log, and the device's clock reading taken at collection time), perf (formatted counters plus both process top-10s). Payloads are depth-safe by construction — flat objects and one paged stream per list, entries at depth ≤3 — so the server's result sanitizer stores them verbatim (pinned by the round-trip fixtures). What's New in the client names the nine checks, the never-read recovery keys, and the snapshot honesty note.
- Manage billing, first-class (My Account). The Billing tab's portal button is now Manage billing: it opens Stripe's secure billing portal (update card, download invoices, cancel — all on Stripe's own pages), and it only renders when billing is actually reachable (Stripe configured + a billing customer on file), so it can never be a dead button — the billing card explains the state instead. The portal endpoint also survives the classic first-run failure: if the Stripe account has no billing-portal configuration yet, the server creates a sensible default configuration through the Stripe API once and retries, instead of failing forever. A genuinely unavailable portal returns an honest "not available yet" error with no Stripe internals leaked, and an account with no billing customer gets a plain "opens after your first payment" message.
- Uptime history on /status. The public status page now shows 30- and 90-day uptime percentages computed only from the page's own incident records (owner-posted incidents, kept since server 2.5.125): time with an open incident counts as down, overlapping incidents are merged so they never double-count. Coverage starts at the earlier of the first incident on record and an
uptime_since tracking marker stamped by this release — where the record history is shorter than the window, the page says "records since <date>" rather than implying a full 90 days exists. The customer site-monitor add-on was inspected and deliberately NOT used: it tracks customers' sites and keeps no history of this service. Zero incidents reads 100.00% over the covered range. - Customer API keys (Business / Custom). Business, Enterprise, and Custom plan accounts can mint read-only API keys in My Account → Security → API access. Keys are SHA-256-hashed at rest with only a short prefix stored for display; the plaintext is shown exactly once, at creation. Two endpoints, scoped server-side to the key owner's account with no way to widen scope:
GET /api/v1/devices and GET /api/v1/detections (Bearer auth, 120 requests/minute per key, GET only — POST answers 404). The plan gate is enforced on the server from the account's live licenses (a fixed-tier account gets 403), revocation is instant (revoked key → 401), and the endpoints are documented on /business and in Help with a copy-paste curl example. - Win-back drip sequence. Ended licenses (canceled / expired / deleted — the 2.5.121 win-back population) now get automated touches at day 1, day 7, and day 21 after the end date, via a new daily
winback-drip job using the existing (owner-editable) win-back email template; touch 2 and 3 vary the subject unless the owner has customized the template. One winback_touches row per (license, touch) is the single source of truth for both paths: the manual admin send now records its touch too, so a manual email covers the scheduled touch for its window and nobody is ever double-sent; a reactivated license leaves the population and the sequence simply stops. Every send lands in the mail log. The jobs viewer lists winback-drip honestly as NOT Run-Now-safe (it sends real email); the owner's manual trigger is the dedicated, audited Customers → Win-back run endpoint. Server-only release; the Windows client stays 2.5.118. Nothing for Austin to configure (SMTP still required for any email, as before).
Version 2.5.175 (admin-only)
- Tech tools are their own admin category. The 29 read-only device checks no longer live inside the install detail page — the admin sidebar has a top-level Tech tools category (same access as the install pages: owner + manager; the support role's page never carries the tab — its markup is emitted only for roles that can use the installs surface — and its API calls stay refused). The category page opens with a device picker: every install with its nickname/name, platform, online state (last heartbeat within 15 minutes) and owner account, a search box filtering by name/nickname, and a checkbox per device. Clicking a device loads that device's full panel — all 29 tool cards with Run buttons (including the Drive check's per-drive target select), live progress bars with estimates, Copy customer update, Cancel buttons, stored results and previous runs. The panel code is the same code, moved: one copy of the renderers (
diagRenderTool / diagRenderResult / diagSummary / diagProgressHtml and friends) now lives in the Tech tools section of the page, still in the served /admin HTML; the data still comes from GET /api/admin/installs/:id/diag, runs still queue through POST /api/admin/commands, the 5-second progress poll still applies while the selected device has a pending run (it now also stops when the selection changes, so pollers never stack), and the section loads independently with error + Retry like every other admin section. Deep link: #tab=techtools&install=<id> preselects the device. - Run one check on several devices at once. With one or more devices ticked, a tool dropdown (all 29 friendly names) and Run on selected devices queue that check on each device through the same queue endpoint (the Drive check runs as all drives in this mode, and the page says so). The outcome is a per-device list — queued, or failed with the reason — with an Open button per device that loads it in the panel; one device's failure never blocks the others. A ticked device that no longer appears in the picker list is reported failed inline instead of being queued (the queue endpoint accepts any id, so posting a stale one would fake a success). Fees behave exactly as single runs: a completed run bills per device when a price is set; errors and cancels never bill.
- Install page keeps a compact card. The install detail page's embedded panel is replaced by a summary card: the latest result line per check that has one (tool — when — one-line summary via the same
diagSummary), any currently pending run noted, and an Open in Tech tools button that deep-links to the category with the device preselected. The service-notice state stays where it already lives, in the fees area — fees did not move. - No endpoint added, no plumbing changed. The picker reuses
GET /api/admin/installs; queueing, progress, cancel, results and fee creation are byte-for-byte the same paths as 2.5.174 — only the page structure moved. The admin guide's Tech tools section now leads with the category + multi-run and keeps the round-by-round notes; the homepage is not rotated (admin-only). Tests: new server/tests/server/v25175-techtools-tab-test.js (nav/category by role — owner + manager see it, support doesn't and its diag/commands calls still 403; picker lists devices with online state and the search filter works; selecting a device renders all 29 cards; a full fake-device flow on the category page — queue → progress → result, including the 2.5.174-fixed Hardware check with the canonical payload — plus cancel from the category page; a 3-device multi-run with one stale id failing inline while the others succeed and completed runs creating fees per device when priced; the install page's compact card; deep-link preselect; served-page structure pins — one panel copy, no full-panel markup on the install page). v25174-diagrender-test.js stays green unchanged (the renderers moved within the same served page's script). Older suites' version-identity pins updated to 2.5.175 per standing practice; the soak baseline is regenerated for 2.5.175 (endpoint count unchanged: no endpoint was added).
Version 2.5.174 (admin-only)
- Tech tools panel fix — renderer/payload key drift, found by a full 29-tool audit. Three admin renderers read payload keys the Windows client has never emitted, so results stored correctly but rendered empty (or partly blank) in the panel: Hardware inventory read
board / gpu / ram (with vram_gb, slot, capacity_gb subfields) while the client has emitted baseboard / gpus / memory (with vram_mb, locator, size_gb) since round 2 (client 2.5.133) — the motherboard, GPU, and RAM sections of every stored hardware result rendered empty; the renderer now reads the emitted keys and also shows the board version and total memory. Startup items read items / source while the client emits entries / location — every result rendered "No startup items found"; it now renders the entries and notes when the found total exceeds the shown list. Windows Update read a last_results object and a per-update caption while the client emits flat last_detect / last_download / last_install, a last_errors object, and per-update description — the last-results line always said "not reported" and descriptions were blank; it now shows the real timestamps, the recorded error codes when any exist, and the descriptions, and the one-line summary counts from installed_count (the client's shown list is capped at 30). The matching one-line summaries for hardware and startup are fixed the same way. In every case the reader was fixed to the shipped client's payload — no client change, and previously stored results render correctly with this server alone. - The audit, not just the three fixes: every one of the 29 tools' admin renderers + summaries was mechanically compared against the payload keys its collector in
src/diag.c actually emits (emitter keys extracted per collector function, reader keys extracted per renderer branch). The other 26 tools match. The customer-facing readers were swept too: the Copy-customer-update sentences and the progress view read only tool names, stage/progress-row fields, and the drive scope (drives[].index / drives[].model, both real emitted keys); the diagnostic-fee auto-create reads only the result envelope (cancelled, error, and diag-sfc's verdict); My Account's service-fees view reads fee-table columns, never payloads. No other drift found. - Tests: new
server/tests/server/v25174-diagrender-test.js makes the audit permanent — it posts one canonical payload per tool (29), shaped key-for-key like the client's emitters in src/diag.c, through the real ingest (POST /api/diag-result → storage sanitizer → GET /api/admin/installs/:id/diag), then evaluates the admin page's own diagRenderResult / diagSummary functions (extracted from the served /admin page) against each read-back result and asserts the rendered HTML and the summary contain the fixture's distinctive values (a board product string, a GPU name, a startup entry, an update description, a printer name, a program name, one per tool) — the assertions that would have caught this drift in round 2. Older suites' version-identity pins updated to 2.5.174 per standing practice; the soak baseline is regenerated for 2.5.174. Server-only: no client binary, no manifest change, no homepage rotation (admin-only).
Version 2.5.173
- Tech tools round 5 (with Client 2.5.138): six more read-only per-device checks, wired in all four places (server
VALID_CMDS, remoteCommandIsValid/diag-cancel names in DIAG_TOOL_NAMES, client parser whitelist + cancel names, both engine dispatchers via diag_is_tool), bare commands only — diag-drive remains the only diag tool that takes an argument. System spec sheet (diag-spec): the ticket card — OS name from Win32_OperatingSystem.Caption (deliberately NOT the registry ProductName, which still says "Windows 10" on Windows 11), version + build, DisplayVersion release, install date from the registry InstallDate DWORD (Unix seconds, converted by a unit-tested pure helper), BIOS manufacturer / version / release date (Win32_BIOS), CPU model + cores + logical processors + socket count (Win32_Processor), motherboard, RAM total + slot count, GPUs, the boot volume, and the boot drive's identity resolved through the drive check's own WMI association chain — the boot_drive object is omitted when unmapped, never guessed. The guide copy states plainly that the sheet compiles the hardware collectors' fields; the only new data is BIOS + OS install date. Browser extensions (diag-extensions): Chrome + Edge per user profile (C:\Users\<user> enumeration, Default + Profile * under each browser's User Data) reading each extension's manifest.json — the name resolved through the extension's _locales\<default_locale>\messages.json with a unit-tested pure chain (__MSG_ key → default locale → en → the raw key → the extension id) — plus Firefox profiles discovered through profiles.ini (unit-tested pure parser) and read from extensions.json (only type == "extension" add-ons; name + creator from defaultLocale). Caps: 8 users, 6 profiles per user per browser, 200 extensions per profile, 8 profiles listed, truncated flags. Names, versions and ids only — never extension content or settings. Hosts & proxy (diag-hosts): every non-comment, non-blank hosts line verbatim (cap 200 lines), plus per-user Internet Settings proxy (ProxyEnable / ProxyServer / AutoConfigURL) read from each loaded hive under HKEY_USERS only, SIDs resolved to account names via ProfileList, the service hive labeled as the system account; profiles not loaded are counted as skipped, never force-loaded (the guide states why: no hive-loading games on a customer PC). Disk error history (diag-diskhist): the 90-day disk-only lens over the same dynamically loaded wevtapi — System channel, levels 1–3 (errors AND warnings), providers disk/Disk/Ntfs/Microsoft-Windows-Ntfs/volmgr filtered in the XPath AND re-checked with a unit-tested pure exact-case-insensitive predicate, grouped by (provider, event id) with counts + newest message, top 25 groups, same 2000-record / 60-second caps as diag-events. The guide states the triangle plainly: diag-events = 7 days everything; diag-diskhist = 90 days disk-only; diag-drive = the disk's state right now. Scheduled tasks (diag-tasks): ALL tasks parsed from the %windir%\System32\Tasks file tree (the diag-startup file approach, extended past logon triggers): full task path, hidden flag (file attribute OR the task's <Hidden> setting), the distinct trigger-type words present, up to three action program paths ("COM handler" when the action lives in the registry), and Microsoft marking (path under \Microsoft\); cap 300, truncated flag, non-Microsoft tasks emitted first; the guide notes the startup check reads the logon-trigger subset of this same tree. Temperatures (diag-temp): ACPI thermal zones (MSAcpi_ThermalZoneTemperature, CurrentTemperature in tenths of a Kelvin converted by a unit-tested pure helper with a sanity range that rejects implausible readings, zone name from the instance path's last segment) PLUS per-drive temperature via the round-4 storage reliability collection, so the tool still answers without ACPI zones. Fee rule, same as the round-4 bootperf precedent: when NEITHER source reports anything the result is error-shaped — {"error":"No temperature sensors are reported by this PC"} — no data, no fee; temperatures are never estimated and never borrowed from another sensor. Round 5's structural fix, stated plainly: the server's result sanitizer caps any single array at 40 entries, so the three list-heavy tools emit their lists as pages (arrays of ≤40-entry arrays: task_pages, ext_pages, line_pages), which is what lets the stated caps (300 tasks, 200 extensions per profile, 200 hosts lines) survive ingest intact — extensions travel as one paged stream with each entry naming its browser profile, because entries nested inside per-profile objects would land past the sanitizer's depth cap and store as empty objects (caught by this release's own round-trip test before ship) — verified by server tests that post more than 40 of each and read back every entry; emission is additionally byte-bounded (~28 KB against the live buffer length) with true count totals and truncated flags whenever any cap bites, and the admin renderers flatten the pages. (Prior rounds' flat arrays silently store-truncate at 40 entries — pre-existing behavior, unchanged by this release.) All six: live progress stages, cancel boundary checks between queries/sources, journaled on the device, friendly names in the server DIAG_FRIENDLY table and the client's balloon table, per-tool admin renderers + summaries + explainer notes, and the standard fee rule — a completed run bills when a price is set; an error (including the temperature no-data result) or a cancel never does. - Copy: ToS §4s and the Privacy Policy diagnostics bullet extended for the six new reports, naming the sensitive fields plainly — the hosts report includes hosts entries verbatim (a hosts file can carry internal network names) and proxy server / auto-config addresses, the extensions report names extension names, and the spec sheet is stated to compile the identity fields the other hardware reports already carry (including the motherboard serial number). /tech-services (EN + ES) lists the six new checks; the homepage "New in" rotates (Server 2.5.173 + Client 2.5.138 + Server 2.5.172 visible; Client 2.5.137 heads the fold, named in the fold summary) and the Technician services card enumerates all twenty-nine checks; the admin guide gains the round-5 paragraph including the overlap triangles, the loaded-hive rule, and the temperature no-data rule.
- Tests:
v25173-diag5-test.js — four-place guards for the six tools, validator matrices (bare accepted, any :arg rejected, diag-cancel accepts the six), per-tool payload round-trips through a fake device (queue → poll → progress → post → admin read-back asserting verbatim round-trip + summaries + renderer content), paging pins (105 tasks, 45 extensions in one profile, 55 hosts lines posted in pages and read back complete — the sanitizer's 40-per-array cap is exactly what the pages defeat), fees for the six (priced completions bill; error, cancelled, and the temperature no-data results do not), Terms/Privacy/tech-services/homepage content pins, and the inline-script regression. Client: diag6-test.c unit-tests the round-5 pure helpers (JSON string extraction with escapes, __MSG_ key parsing, locale lookup incl. case-insensitive keys and the missing-locale fallback, profiles.ini section walking, the Firefox addons span walker incl. string-brace traps, hosts line filtering, task trigger/action/hidden/Microsoft extraction, ACPI tenths-of-Kelvin conversion incl. sanity rejection, the ACPI zone-name cut, the disk-provider set, the Unix-seconds install date) and diag6-test.sh pins the wiring (per-tool mentions in diag.c/commands.c, collector sources — Caption not ProductName, loaded hives only, the Tasks tree, the reliability counter reuse — notice strings, 'safe to restart' absent from every client source, version floor, What's New entry); parse cases for the six join cmd-test.c. Older suites' stale rotation/version pins (v25169, v25170, v25171, v25172) updated to the 2.5.173 state per the standing practice, and the soak baseline is regenerated for 2.5.173.
Version 2.5.138
- Tech tools round 5 collectors (with Server 2.5.173):
src/diag.c gains the six collectors above on the established patterns — the spec sheet reuses the hardware check's board/GPU/memory row callbacks and the drive check's association-chain tables verbatim; extensions reads browser metadata files with bounded reads (manifests ≤256 KB, extensions.json ≤1 MB); hosts reads the hosts file plus loaded-hive registry values only (no hive is ever loaded by this client); disk history reuses the dynamic wevtapi loader and the events grouping table; tasks walks the task-file tree with the startup check's bounds; temperature reuses the round-4 storage callbacks for drive temperatures. Pure, unit-tested helpers live in a new DIAG5-TESTABLE block (span-bounded JSON string extraction, __MSG_ resolution, locale lookup, profiles.ini + extensions.json walkers, hosts line filtering, task XML trigger/action/hidden/Microsoft extraction, Kelvin conversion, zone naming, the disk-provider predicate, the install-date conversion). Every collector checks the cancel flag between its queries/sources, reports progress stages, and journals its run like rounds 1–4. Nothing any of them reads is written back: no hosts entry edited, no proxy changed, no extension touched, no task modified, no sensor configured. - What's New in the Feature Guide covers round 5 in the standing single-line format, including the spec-sheet compilation note, the signed-in-only proxy rule, and the no-sensors temperature note.
- Linux devices report as Linux (platform normalization). The Linux daemon has always identified itself to the server as
"LX" in its register and heartbeat requests, but the server's platform validator only accepted the full names (windows / linux / android) and silently filed anything else under Windows. Consequences: every Linux install showed as a Windows PC in the admin console (each heartbeat rewrote the stored platform to windows), and a Linux daemon registering directly with a Linux-scoped license key was refused with the backwards message "This Linux license key cannot be used on Windows." validPlatform() now normalizes the value case-insensitively through an alias map of the strings the shipped clients actually send (lx → linux; the Windows client sends windows and the Android app sends android, already canonical), and the same normalization is applied at checkout, in client bug reports, and in admin bulk-key creation. Genuinely unknown values keep the previous default. Already-flipped Linux installs heal themselves: the daemon reports its platform on every heartbeat and the heartbeat update writes it, so each affected install flips back to linux on its first check-in after this deploy — no database repair needed and no daemon update required. Platform-scope refusals now fire only for real cross-platform misuse and name the platforms in the right direction. - The download page's hardcoded Linux fallback version label now reads 1.23.1 (the Linux client release shipping alongside this fix). Regression coverage in
server/tests/server/v25138-platform-normalize-test.js (LX / lx / mixed-case registrations stored as linux, Windows/Android registrations unchanged, both refusal directions with correct messages, a pre-flipped install healing on its next heartbeat, and the admin installs API view). Server-only release; the Windows client stays 2.5.117. Nothing for Austin to configure.
Version 2.5.172
- Tech tools round 4 (with Client 2.5.137): six more read-only per-device checks, wired in all four places (server
VALID_CMDS, remoteCommandIsValid/diag-cancel names in DIAG_TOOL_NAMES, client parser whitelist + cancel names, both engine dispatchers via diag_is_tool), bare commands only — diag-drive remains the only diag tool that takes an argument. Drive reliability (diag-ssd): per physical disk from root\Microsoft\Windows\Storage — MSFT_PhysicalDisk joined to MSFT_StorageReliabilityCounter by DeviceId (unit-tested pure join): model, serial, media type (HDD/SSD/SCM where the provider names it), wear percent, power-on hours, temperature, and read/write error totals. Every field is emitted only when the provider returned it — most HDDs and many SSDs report some or none, and the panel reads "not reported" per field, never a synthesized value; the renderer shows a wear bar like the battery one. Network adapters (diag-adapters): Win32_NetworkAdapter (PhysicalAdapter filter, with an unfiltered fallback if the provider rejects it) joined to Win32_NetworkAdapterConfiguration by Index (IP addresses, gateway, DNS — string arrays flattened, capped at 4 entries each) and to Win32_PnPSignedDriver (DeviceClass 'NET', joined on PNPDeviceID == DeviceID) for driver version + driver date; link speed from the adapter's Speed in Mbps where reported (often absent on Wi-Fi). Plus the current Wi-Fi connection via the client's existing WlanAPI code: SSID, signal percent, channel (via wlan_intf_opcode_channel_number where it succeeds), PHY type in words, and a band word claimed only when the PHY type pins it (802.11a/ac/ad → 5 GHz, b/g → 2.4 GHz; n/ax span both bands and get no band claim). Restore points (diag-restore): root\default SystemRestore — sequence number, creation time, description, and the type in plain words (program installed / program removed / system checkpoint / device driver installed / settings change / cancelled operation / "type N"), sorted newest-first on the client, 30 shown with the true total; shadow-storage usage is parsed verbatim out of vssadmin list shadowstorage output (strict single-line parse, 30-second bound) and is simply absent — "not reported" — when vssadmin cannot run (it needs elevation) or prints nothing parseable; never estimated. Boot performance (diag-bootperf): Event 100 records from Microsoft-Windows-Diagnostics-Performance/Operational via the same dynamically loaded wevtapi as diag-events/diag-crashes — boot, main-path, and post-boot durations in milliseconds extracted from the events' named <Data> elements by a unit-tested pure parser — last 10 boots newest-first, plus a count of degradation events (101–199) at or after the oldest collected boot (client-side UTC string compare; a 90-day XPath window when the log holds no boots). Fee rule, decided and pinned: when the channel is absent/disabled or the query fails, the result is error-shaped — {"error":"boot performance log not available on this PC"} — because no data was produced, so no diagnostic fee is created (the existing maybeAutoDiagFee error refusal covers it); a channel that answers with zero Event 100s is a completed, billable run with an empty boot list. Activation status (diag-license): SoftwareLicensingProduct rows, kept only when they carry a partial product key (the class lists hundreds of keyless entries) — product name, the channel word parsed from the Description by a unit-tested pure helper (KMS → "Volume (KMS)" wins over VOLUME → "Volume", then RETAIL, then OEM), LicenseStatus in plain words (Licensed / In grace period / Notification (not activated) / Unlicensed / "status N"), and grace days from GracePeriodRemaining where a grace status reports them; covers Windows and Office alike. Privacy rule, enforced in the collector: the partial key is read into a scratch buffer only to test non-emptiness — it is never copied into the payload, never logged; source guards + a server payload guard pin that no key field exists anywhere in the report. USB storage history (diag-usbhist): the registry's HKLM\SYSTEM\CurrentControlSet\Enum\USBSTOR tree — per device class + per instance: friendly name (the instance's FriendlyName value where present, else the class descriptor words via a unit-tested pure parser — "Disk&Ven_SanDisk&Prod_Ultra&Rev_1.00" → "SanDisk Ultra"), the serial (the instance-id segment, verbatim), and the instance key's last-write time via RegQueryInfoKey converted by a unit-tested pure FILETIME→date helper and reported as "last seen (from registry)", labeled approximate in the renderer and the guide. Names + serials only — the registry holds no file or content data for these devices, and the explainer says so. All six: live progress stages, cancel boundary checks between queries, journaled on the device, friendly names in the server DIAG_FRIENDLY table and the client's balloon table, per-tool admin renderers + summaries + explainer notes, and the standard fee rule — a completed run bills when a price is set; an error (including the bootperf unavailable-log result) or a cancel never does. - Copy: ToS §4s and the Privacy Policy diagnostics bullet extended for the six new reports, naming the sensitive fields plainly — the adapters report includes MAC addresses and the current network name (SSID), the USB history report includes storage-device serial numbers, and the activation report states explicitly that no product keys are collected (partial keys are used only to filter, never stored or sent). /tech-services (EN + ES) lists the six new checks; the homepage "New in" rotates (Server 2.5.172 + Client 2.5.137 + Server 2.5.171 visible; Client 2.5.136 heads the fold, named in the fold summary) and the Technician services card enumerates all twenty-three checks; the admin guide gains the round-4 paragraph including the boot-performance availability note.
- Tests:
v25172-diag4-test.js — four-place guards for the six tools, validator matrices (bare accepted, any :arg rejected, diag-cancel accepts the six), per-tool payload round-trips through a fake device (queue → poll → progress → post → admin read-back asserting verbatim round-trip + summaries + renderer content), the bootperf unavailable-log → error → no-fee pin, the diag-license no-key payload guard, fees for the six (priced completions bill, error and cancelled do not), a cancelled run recording no duration, Terms/Privacy/tech-services/homepage content pins, and the inline-script regression. Client: diag5-test.c unit-tests the round-4 pure helpers (reliability join by device id, media words, restore-point type words, the vssadmin used-line parse, Event 100 named-data extraction from canned XML incl. degradation-window inputs, license status/channel mapping, USBSTOR class-name parsing, FILETIME→date vectors, PHY/band honesty) and diag5-test.sh pins the wiring (per-tool mentions in diag.c/commands.c, notice strings, 'safe to restart' absent from every client source, version floor, What's New entry); parse cases for the six join cmd-test.c. Older suites' stale rotation/version pins (v25169, v25170, v25171) updated to the 2.5.172 state per the standing practice, and the soak baseline is regenerated for 2.5.172.
Version 2.5.137
- Tech tools round 4 collectors (with Server 2.5.172):
src/diag.c gains the six collectors above on the established patterns — Storage-namespace WMI row callbacks for drive reliability, CIMV2 joins for adapters (plus a new string-array property reader and the WlanAPI current-connection read modeled on wifi_current() in paid5.c), the SystemRestore class in root\default plus a bounded vssadmin capture for restore points, the dynamic wevtapi loader for boot performance, SoftwareLicensingProduct scanning for activation (the partial key never leaves a stack scratch buffer), and a bounded registry walk for USB history. Pure, unit-tested helpers live in a new DIAG4-TESTABLE block (media words, reliability join, restore type words, vssadmin line parse, boot-record XML parse, license status/channel words, USBSTOR name parse, FILETIME→date, PHY/band words). Every collector checks the cancel flag between its queries, reports progress stages (bootperf posts a second stage after the boot scan, before counting degradations), and journals its run like rounds 1–3. Nothing any of them reads is written back: no drive is touched, no adapter reconfigured, no restore point created or deleted, no license changed, no registry value written. - What's New in the Feature Guide covers round 4 in the standing single-line format, including the no-product-keys note and the boot-log availability note.
- App-firewall blocklist UI, completed (Fleet). The fleet policy's
blocked_apps field — already delivered to clients in the heartbeat as policy_blocked_apps and already enforced by the Windows client (blocked_apps_parse in cloud.c / the guard watcher) — finally gets a proper admin surface: Protection → App blocklist lists each blocked app as its own row with add/remove, backed by the SAME policy field and written through the same writeFleetPolicy path as the policy form's raw "Blocked applications" textarea, so the two surfaces can never diverge (the textarea is also re-synced whenever the list changes and whenever the policy loads). Validation mirrors the client's parser contract: ;-separated exe names or full paths, trimmed, max 260 chars per entry, max 64 entries; entries containing ;, quotes or control characters are rejected with a 400, and duplicates (case-insensitive) are refused. New endpoints: GET /api/admin/policy/blocked-apps, POST /api/admin/policy/blocked-apps/add, POST /api/admin/policy/blocked-apps/remove — adds and removes are audit-logged (policy.blocked_apps.add / policy.blocked_apps.remove). No client change was needed — the client hook was verified present — so the Windows client stays 2.5.116. - Helper admin accounts. New
admin_users table (email, scrypt password hash via the existing hashPassword/verifyPassword helpers, role support|manager, disabled flag, created_at, last_login). Helpers sign in at the SAME /admin-login — supplying an email selects the helper path; leaving it blank keeps the owner password login working exactly as before — and a bad helper login returns the exact same generic wrong password error as a bad owner login (no account-existence hints), under the same brute-force lockout. Sessions carry the helper identity + role. Enforcement is server-side on every admin endpoint, inside adminAuth: Support gets an explicit allowlist only (customers READ — list/detail/timeline/charges; customer notes GET/POST/DELETE; support inbox read/archive/reply + contact list; expiring-soon list + reminder send; win-back list + send; own-password change; whoami/logout) and everything else returns 403 — explicitly including refunds, backup, feature flags, config viewer, exports, helper management, policy/blocklist edits, broadcast and promos. Managers get everything EXCEPT an owner-only denylist: Stripe refunds (/api/admin/accounts/:id/refund), backup downloads (/api/admin/backup, /api/admin/backup/download, /api/admin/backups/download), feature flags (/api/admin/feature-flags), and helper management (/api/admin/helpers*, additionally behind an explicit requireAdminOwner middleware). Unlisted endpoints default to owner+manager. Every helper action audit-logs under the helper's email (an AsyncLocalStorage actor flows from adminAuth into auditLog's default actor, and adminEmailOf returns the helper email for DB audit rows). Disabling a helper drops their live sessions immediately AND is re-checked against the DB on every request, so an existing session is rejected on its next call. Owner UI at System → Helper admins: create (email + role → the server generates a one-time temp password shown EXACTLY once, in the create response only), list, disable/enable, delete; helpers change their own password from the same section (POST /api/admin/session/password). The full enforcement map is documented in the admin guide. - Scheduled fleet scans (server half). The fleet policy gains
scan_schedule: { enabled, type: 'quick'|'full', days: [0-6] (0 = Sunday), time: 'HH:MM' }, sanitized beside the other policy sanitizers by the new sanitizeScanSchedule — a bad type falls back to full, a bad time (e.g. 25:99) to 02:00, out-of-range or duplicate days are dropped, and the rest of the policy is left intact. Editable controls live in Protection → Policy → Updates & patching → Scheduled fleet scans (wired into the fleet policy form only), and the sanitized schedule is delivered in the existing heartbeat policy payload as policy_scan_schedule, next to policy_blocked_apps. Client consumption lands in a later client release; this release is save + sanitize + deliver only. - Homepage "New in" rotation keeps Server 2.5.137 / Server 2.5.136 / Client 2.5.116 visible, with Server 2.5.135 folded into the single "See older updates" collapsible (label now runs to Server 2.5.135). Help & FAQ gains entries for the fleet app blocklist + scan schedule and for helper admin sign-ins; the admin guide documents all three features including the full role-enforcement map. Regression coverage in
server/tests/server/v25137-batch8r2-test.js. Server-only release; the Windows client stays 2.5.116. Nothing for Austin to configure.
Version 2.5.171
- Tech tools round 3 (with Client 2.5.136): six more read-only per-device checks, wired in all four places (server
VALID_CMDS, remoteCommandIsValid/diag-cancel names, client parser whitelist + cancel names, both engine dispatchers via diag_is_tool), bare commands only — diag-drive remains the only diag tool that takes an argument. Printers (diag-printer): Win32_Printer name/default/driver plus a plain-words status composed from PrinterStatus + DetectedErrorState + WorkOffline (real problems outrank activity states), and a per-printer queued-job count from Win32_PrintJob (jobs belong by the "Printer, <id>" name rule). Installed programs (diag-programs): the Uninstall registry keys (HKLM 64-bit view, HKLM WOW6432Node view, HKCU) — DisplayName/Publisher/DisplayVersion/InstallDate (YYYYMMDD normalized to ISO)/EstimatedSize; entries with a ParentKeyName (patches) and SystemComponent entries are skipped, identical (name, version, publisher) deduped across hives, capped at 500 with a truncated flag, sorted by name. Problem devices (diag-devices): Win32_PnPEntity where ConfigManagerErrorCode != 0, with a plain-words map for the known codes (1/10/12/18/19/28/31/43) and "error code N" for the rest — zero problem devices is a first-class clean result. Services not running (diag-services): Win32_Service with StartMode Auto and State != Running (name, display name, state, exit code), capped at 50, plus auto_total/auto_running so the panel shows proportion. User accounts (diag-users): Win32_UserAccount LocalAccount=true (name, Disabled, PasswordRequired), Administrators membership via Win32_GroupUser matched on the group's locale-independent SID S-1-5-32-544, last logon from Win32_NetworkLoginProfile normalized where present and omitted — the panel says "not reported" — where it is not (never derived). Crash history (diag-crashes): System-log bugcheck events (Microsoft-Windows-WER-SystemErrorReporting, Event 1001) over 90 days via the same dynamically loaded wevtapi as diag-events, stop code parsed from the "The bugcheck was: 0x…" token (unparseable reports shown raw, capped), plus the C:\Windows\Minidump inventory (names, dates, sizes only, newest first, cap 20; an absent folder is reported as absent). All six: live progress stages, cancel boundary checks, journaled on the device, friendly names in the server DIAG_FRIENDLY table and the client's balloon table, per-tool admin renderers + summaries, and the standard fee rule — a completed run bills when a price is set, an error or a cancel never does. - Repairs gain an open/completed state + the device service notice:
service_fees.repair_status (migration; pre-existing repairs default to 'completed', new records start 'open') is separate from the billing status. Recording a repair sets the install's service notice ON (installs.svc_notice_on/since/repair, migration) unless the admin unticks the new Show service notice on the device checkbox (default on) — POST /api/admin/installs/:id/repairs accepts notify_device. Complete repair (POST /api/admin/installs/:id/repairs/:fid/complete) closes the record and turns the notice off when that repair owns it; Clear service notice (POST /api/admin/installs/:id/svc-notice/clear) clears any stuck notice without touching the repair. The heartbeat response carries svc_notice on every beat so devices converge. The install page's fees section shows an ON notice prominently with its age in hours/days (there is deliberately no automatic expiry — a mailed-part repair can span weeks — so the panel is the safety net), repair rows show an open/completed badge and the Complete button, and the Customers → Service fees table reuses the same row renderer. All transitions are audit-logged (svc_notice.on/off, service_fee.complete). - Copy: ToS §4s and the Privacy Policy diagnostics bullet extended for the six new reports — the Privacy text states plainly that the user-accounts report includes the device's local account names — plus the on-device notice behavior; /tech-services (EN + ES) lists the six new checks and explains the notice; admin guide gains the round-3 + notice-lifecycle paragraph and the Service fees paragraph is corrected for the notice; homepage "New in" rotates (2.5.171 + Client 2.5.136 visible, 2.5.169 folds into older updates) and the Technician services card lists all seventeen checks.
- Tests:
v25171-diag3-test.js — four-place guards for the six tools, per-tool payload round-trips through a fake device (queue → poll → post → admin GET), renderer + summary content pins against the served admin page, the svc_notice lifecycle (record → on / unticked → stays off / heartbeat carries it / complete → off / manual clear / fees GET state + age), fees for the new tools (priced completion bills, error and cancelled do not), Terms/Privacy/tech-services content pins, and the inline-script regression. Client: diag4-test.c unit-tests the round-3 pure helpers (printer status precedence, ConfigManager code words, program dedupe/cap/sort/date normalization, print-job ownership, bugcheck stop-code parsing, logon normalization, GroupUser name parsing) and diag4-test.sh pins the notice wiring (balloon strings, banner reuse, and 'safe to restart' absent from every client string).
Version 2.5.136
- Tech tools round 3 collectors (with Server 2.5.171):
src/diag.c gains the six collectors above on the established patterns — WMI row callbacks for printer/devices/services/users, registry-view enumeration for programs (reusing the existing diag_reg_* helpers), and the dynamic wevtapi loader for crashes — with pure, unit-tested helpers in a new DIAG3-TESTABLE block (printer status words, problem-code words, program dedupe table, job ownership, stop-code parse, logon normalization). diag_prop_u64 now also accepts the VT_I2/VT_UI2/VT_UI1 variants printer and device codes arrive as. Every collector checks the cancel flag between its queries, reports progress stages, and journals its run like rounds 1–2. Nothing any of them reads is written back: no printer is touched, no service started, no account changed, no dump file opened. - In-progress notice (Part B): when a check starts, the customer gets one tray balloon — "<Check name> - a technician is running this check on your PC. Please don't turn off or restart until it finishes." — and a persistent amber banner at the top of the main window (a new
IDC_SVC_BANNER static created beside the expired-license banner, sharing its amber paint, repositioned under it when both show, refreshed by the same 5-second refresh_status loop; non-modal, SWP_NOACTIVATE, never steals focus). The banner names the running check (or "technician checks" when several run). When the last active run ends — done, errored, or cancelled — the banner clears and one balloon follows: "The check on your PC has finished - please check with your technician for the next step before you restart." An owed-flag consumed via InterlockedExchange guarantees exactly one completion balloon even for overlapping runs. The phrase "safe to restart" appears nowhere in the client (pinned by a source guard): the completion wording always routes the customer to the technician for the next step. - Repair service notice: the heartbeat's
svc_notice is applied in cloud.c on every beat — state persists in config.ini (svc_notice_on/svc_notice_since) so the banner ("This PC is being serviced by your technician - please don't turn it off or restart." + "service started <date>") returns after a restart for as long as the repair stays open. One balloon fires when the notice turns on; when the technician completes or clears it, one balloon: "Your technician has marked the service complete - please check with them for the next step before you restart." When a check and the service notice are active together, the single banner shows both facts, check line first. - What's New in the Feature Guide covers round 3 and the notices in the standing single-line format.
- Onboarding checklist (My Account).
/account gains a "Get fully protected" card whose steps are computed ONLY from real server-side state for the signed-in account: account created (the always-done anchor), email verified (accounts.email_verified), first device installed (an installs row under one of the account's licenses), a device nicknamed (the new installs.label set on any of the account's installs), and two-factor authentication enabled (an enabled account_2fa row). Nothing client-reported and nothing unverifiable counts. Each unfinished step links to where it is done (the download page, the device list, the Security tab); when every step is done the card collapses to a short "You're fully set up" line. - Device nicknames (
installs.label). A new owner-set label column on installs (migration-guard pattern, distinct from the client-synced display_name). Owners set/rename/clear it from the device rows on My Account (POST /api/account/installs/label — ownership enforced server-side via the license join; a non-owner gets 404), and admins from the install detail view (POST /api/admin/install/label, audit-logged as install.label, following the existing rename audit pattern). The label is displayed with hostname fallback everywhere the device appears: the account device list, the admin installs list and install detail, the customer timeline, the protection summary, the Year in Protection report, shared (family) license device lists, and admin global search. CRITICAL invariant: heartbeat and register paths update hostname/display_name but NEVER write label, so a nickname survives reinstalls and renames on the PC. - Invoice downloads (Billing tab).
GET /api/account/billing now returns each invoice's Stripe hosted_invoice_url and invoice_pdf alongside the existing date/amount/status, read through a brief 60-second per-customer server-side cache (the promo-code cache pattern). Invoices are listed STRICTLY for the signed-in account's own Stripe customer id (taken from its licenses). The response also carries stripe_configured / has_customer / message, and the Billing tab says plainly when Stripe is not configured or no billing customer is on file instead of showing an empty table; the rendered rows link out to the hosted invoice page and the PDF. - Download my data, completed. The existing
GET /api/account/export now includes, in addition to the profile / licenses / installs / sessions / alert rules / leak checks it already carried: detections recorded against the account's own devices, family memberships (owned memberships, memberships where the account is the member, and invites sent — the 2.5.134 tables), and subscription + referral state (per-license Stripe linkage, the account's referral code, referrals as referrer and as referred). License keys are included in full — it is the customer's own data behind their login; password hashes and 2FA secrets remain excluded as before. Every query is scoped by the session account id, so another account's rows are structurally impossible to include. Exports are rate-limited to one per hour per account (429 with a plain message) and each export is recorded in the new data_exports table plus the existing audit entry. - Spanish round 2. The 2.5.125 language mechanism (
?lang=es + the 30-day aelang cookie) now also covers /account, /download, /login, and /register: their static labels and copy render in natural Spanish (Mi cuenta, Descargar, Iniciar sesión, Crear cuenta, the account tabs and section headings, the install steps, the free-account and expired-license download cards), while ALL dynamic data — names, keys, amounts, versions, hashes — is identical to English. Legal text stays English with the same Spanish note as round 1. The new checklist, nickname prompts, and billing states render in Spanish on those pages too. - Homepage "New in" rotation keeps Server 2.5.136 / Client 2.5.116 / Server 2.5.135 visible, with Server 2.5.134 folded into the single "See older updates" collapsible. Help & FAQ gains entries for the checklist, device nicknames, invoice links, and the completed data export. Regression coverage in
server/tests/server/v25136-batch8r1-test.js. Server-only release; the Windows client stays 2.5.116. Nothing for Austin to configure.
Version 2.5.170
- Drive identity + per-drive testing (with Client 2.5.135): the Hard drive check's result now identifies each physical drive unambiguously — index (PhysicalDrive N, the WMI
Index), model, serial, size, interface type (InterfaceType), SMART failure prediction, its health line, and the volumes that live on it (letters with free space + dirty flag). The volume→drive mapping comes from the real WMI associations Win32_DiskDriveToDiskPartition → Win32_LogicalDiskToPartition; a volume whose mapping cannot be resolved is reported under unmapped_volumes and rendered under "Volumes not assigned to a drive" — never guessed onto a drive. Results from older clients (no per-drive index/volumes) render exactly as before. - Targeting contract:
diag-drive gains the suite's one optional argument — bare or diag-drive:all = every drive; diag-drive:<n> = only PhysicalDrive n. The exception is enforced identically in all four places (server VALID_CMDS, remoteCommandIsValid, client cmd_parse_one, and the client's diag_is_tool): canonical digits only, no leading zeros, ≤ 999; every other diag tool stays bare-only, and diag-sfc:force-style forms stay rejected. A targeted run collects the same facts scoped to that drive (its volumes only) and carries scope {mode:'one', drive_index:n} vs {mode:'all'}; the progress object carries the same scope — resolved with the drive's model from the newest stored inventory — and the copy-customer-update sentence names it ("a drive health check on Drive 1 (Samsung SSD 980 1TB)"). If drive n is not present, the result is an honest error ("Drive n is not present — this PC has K physical drives") and, per the fees rule, creates no fee. - Admin drive picker: the Drive check card gains a Test select — "All drives" plus one entry per drive from the device's most recent drive inventory (from the latest diag-drive result), each reading "Drive 0 — Samsung SSD 980 1TB · C:, D:" so the admin picks by identity, not number alone. When only a health snapshot exists, the fallback is labeled as volume-only data (which cannot say which drive a volume lives on); with no inventory at all, only "All drives" is offered, with the note "Run Check all drives once to enable per-drive tests."
- Cancel for every Tech tools check: every pending/running card gains a Cancel check button backed by
POST /api/admin/installs/:id/diag-cancel (adminAuth, audit-logged). Queued runs are flagged in the command store and the delivery path (takeCommands) skips cancelled entries — a queued-cancelled check is never delivered, pinned by test; its card shows a neutral Cancelled terminal state and the customer sentence becomes the cancelled variant ("We have cancelled the check on your computer — no further action is needed."). Running runs record cancel_requested_at and a diag-cancel:<tool> command is queued for the device. Cancel targeting, stated plainly: JSON-store command ids are positional per poll (900000000+idx), so the server cannot know the id a run was delivered under — the endpoint therefore targets by tool name; the diag-cancel:<digits> form (a real command id) is implemented and validated on both ends and honored by the client registry for runs that carry a real id (e.g. the SQL device_commands path). All diag bookkeeping (pending state, newest entry, result matching) now matches command entries by base name so targeted forms resolve against their tool. - Truthful outcomes: a device that stops posts
{cancelled:true, stage_reached} instead of partial data; the ingest stamps the command entry cancelled (not done), renders a neutral Cancelled card, and records no run duration (time-to-cancel would skew the time estimates). A real result that lands after a cancel request is annotated after_cancel_requested (badge on the latest run, note in Previous runs) rather than hidden or mislabeled. Fees: maybeAutoDiagFee refuses cancelled:true results — error / SFC failed / SFC timeout / cancelled → no fee, and a queued-cancelled run never runs and never bills; a run that completes despite a cancel request bills normally, as the annotation records. - Tests:
v25170-diag-drive-cancel-test.js — validator accept/reject matrices for both new arg forms (+ the other-tools-stay-bare pins), four-place source guards, and a fake-device live flow (targeted queue → poll → scoped result + fee; unknown index → error stored, no fee; cancel-queued → never delivered + cancelled progress state + cancelled sentence verbatim; cancel-running → diag-cancel:diag-events delivered, cancelled report clears pending, no fee; result-after-cancel → annotation + fee; three cancelled runs leave the estimate source at "guide"; cancel endpoint 400/404/auth shapes). Client pure logic (arg validators, WMI reference parsing, volume→drive mapping incl. interleaved partitions and the unmapped cases) is unit-tested in diag2-test.c; the parse matrix lives in cmd-test.c; wiring guards in diag3-test.sh; a cancelled-no-fee case joins v25167-fees-test.js.
Version 2.5.135
- Drive check identifies and targets drives (with Server 2.5.170):
diag_run_drive reads Index, DeviceID and InterfaceType with the existing model/serial/size/status fields, resolves each volume to its physical drive through the WMI associations (drive → partitions → logical disks), and emits the new shape: scope + per-drive objects carrying index, interface and only their own volumes, plus unmapped_volumes for anything the associations could not place. diag-drive:<n> collects only PhysicalDrive n with the same facts; an absent index returns the honest "not present" error naming how many physical drives the PC has. The reference parsing and letter→drive resolution are pure C in the DIAG2-TESTABLE block (unit-tested on canned association data, including partitions interleaved across two drives and letters with no drive association). - Cancellation: a small running-check registry (base tool, command id, cancel flag; SRWLOCK, statically initialized) is registered by the worker thread before collection.
diag-cancel:<tool> / diag-cancel:<command id> — parsed and validated in cmd_parse_one and dispatched in both dispatchers (commands_dispatch, engine_run_command) — sets the flag on matching runs; no match is a logged no-op. Collectors poll the flag at their natural boundaries: diag-space between top-level folders and inside the walk cap check, diag-events inside the channel loop and between channels, diag-sfc at each 30-second heartbeat slice (cancelling terminates sfc through the existing timeout-termination path), diag-drive after each WMI stage and between drives, diag-startup between source groups and before the publisher pass, diag-hardware between its WMI queries, diag-winupdate after the update-list query, diag-health right after the 1-second CPU sample. Effectively atomic (start-check only, stated plainly): diag-battery, diag-net, diag-posture — single snapshot reads that finish in seconds; for them a cancel lands before the run starts or after it posts. On abort the worker discards whatever was collected and posts {"cancelled":true, "stage_reached": ...} (plus the drive scope for diag-drive) instead, journals diag_cancelled on-device, and acks the outcome cancelled; results always post under the base tool name so targeted runs file under their tool.
- "Update available" notice (admin). The server now compares its running
RELEASE_VERSION with the newest Server version published in the README on GitHub, reusing the changelog machinery (fetchRemoteReadme — now honoring an AE_README_URL env override so tests can stub the fetch; production uses the same raw GitHub URL as always) and parsing the first ## New in Server X.Y.Z header. One mechanism only: the check runs when an admin asks (new GET /api/admin/update-check, adminAuth) and is cached for 6 hours — the cache stamp is taken before the fetch, so a failed check also waits out the TTL instead of hammering GitHub. When GitHub's newest server version is strictly newer, the admin console shows a banner: "Server <X> is pushed and waiting — redeploy Render to update" (with the running version named), and GET /api/admin/server-health gains latest_version / update_available, rendered in System → Server health as a running-vs-latest row. When the fetch fails, times out (5s race so a hung connection can never stall the admin page), or the versions are equal, the UI shows nothing — no banner, no error, no stale claim. - View as customer (admin impersonation). From the customer detail in Admin → Customers, View as customer (
POST /api/admin/accounts/:id/impersonate, adminAuth + requireOwner, browser admin session required — Basic-auth API clients have no session to carry the view) sets a flag on the admin session entry; getAccount() then resolves customer-facing rendering to that account, so /account and every customer page show their licenses, devices and data. Every page rendered during the view carries a persistent banner: "Viewing as <email> — read-only · Exit". Read-only is enforced at the request layer: a gate middleware refuses every non-GET request with 403 while the flag is set — customer action endpoints and admin POSTs alike — with the single exception of the Exit action (POST /api/admin/impersonate/stop). Starting and stopping write audit entries (impersonation.start / impersonation.stop, target account id + email). Exit clears only the flag: the underlying admin session survives, /admin works immediately without re-login, and any customer session in the same browser is untouched. A GET /api/admin/impersonate status endpoint backs the UI state. If the target account is deleted mid-view, the flag drops silently and normal resolution resumes. - Email template editor (Customers → Email templates). New table
email_templates(key, subject, body, updated_at, updated_by) holds OPTIONAL admin overrides for five customer emails: expiry_warning (expiring-soon reminder, renew + past-due variants), trial_warning (the same reminder's trial variant and the trial-nurture job — one template, both send sites), winback (win-back send), gift (Stripe gift delivery in the checkout webhook), and verification (sendVerifyEmail). Each send site keeps building its current subject/body as the fallback and passes it through applyEmailTemplate(key, vars, fallback): with no override row the fallback is returned untouched, so behavior is byte-identical until a template is edited. Substitution is plain {{placeholder}} replacement using only the variables the send site already supplies (enumerated per template: plan / key / date / link for the warnings; + account_link for win-back; plan / key / register_link / download_link for gift; link / email for verification); unknown placeholders render as empty strings and rendering never throws — a malformed stored template still sends. Admin API: GET /api/admin/email-templates (each template with its placeholders, sample data, current effective text and built-in default), POST /api/admin/email-templates (save override), POST /api/admin/email-templates/preview (render with sample data, no save), POST /api/admin/email-templates/reset (delete the override, restoring the default). The editor section follows the independent-loader pattern (its own error + Retry), and every save/reset is audit-logged (email_template.save / email_template.reset). Note the scheduled admin expiry digest to the owner inbox is an internal notification, not a customer email, and is not templated. - Regenerate license key (admin, per-license in the customer view).
POST /api/admin/licenses/regenerate (adminAuth + requireOwner) mints a fresh key with the existing newLicenseKey() helper (same prefix as the old key) and swaps it onto the SAME license row inside a transaction: plan/tier, current_period_end, status, deleted_at, Stripe customer/subscription linkage, account and reseller scope are all untouched, and installs.license_key + family_invites.license_key are re-pointed at the new key so devices and pending invites stay attached. The old key is tombstoned in the new retired_license_keys(old_key, license_id, retired_at) table: POST /api/register checks the tombstone before any account resolution and refuses with the exact existing {ok:false, error:'invalid license key'} shape, and redemption (POST /api/account/redeem-key) finds no license row and answers its existing "That license key was not found." — the old key is dead for NEW registrations and redemptions while existing installs (which authenticate by install_id + token) keep heartbeating untouched. Deleted licenses refuse regeneration with a 400. The admin UI action carries a confirm that states exactly this; the audit entry (license.regenerate) records the license id and the keys' last 4 characters only — never a full key. Help & FAQ's "My license key is rejected" answer now explains the regenerated-key case. - Homepage "New in" rotation keeps Server 2.5.135 / Server 2.5.134 / Server 2.5.133 visible with Client 2.5.115 folded into the single "See older updates" collapsible alongside 2.5.132 and older. The admin guide gains a paragraph covering all four tools. Regression coverage in
server/tests/server/v25135-batch7r2-test.js (update-check ahead/equal/fetch-failure against a stubbed README source across three boots; the full impersonation flow — data isolation, banner, POST refusal with no state change, exit, session survival, audit rows; template default/edit/preview/reset/unknown-placeholder round trips against an SMTP capture via the expiring-soon reminder, win-back and verification sends; regeneration — old key refused at register + redeem with the existing error shapes, new key registers, pre-existing install heartbeat stays 200, license fields unchanged, audit row with last-4s only). Server-only release; the Windows client stays 2.5.115. Nothing for Austin to configure.
Version 2.5.169
- Tech tools live progress + estimated time (with Client 2.5.134): while a check runs on a device, its card on the install page now shows a live progress bar — the stage in plain words (starting, the check's own stage phrase, sending the report), how long it has been running, and the estimated time left — so the technician can keep the customer informed instead of staring at "Running". Devices report real stage boundaries over a new device-authenticated ingest (
POST /api/diag-progress, same auth as the result channel): started when collection begins, collecting at tool-defined points (each event-log channel for Recent system errors; per-folder for Disk space), a still-running heartbeat every ~30 seconds for System file verification, and uploading just before the report post. Progress posts are best-effort — a failed post never fails or delays the check itself. - Honest percentages, honest estimates: a percentage is shown only where the device genuinely knows one — Disk space (folders finished out of folders found on the system drive). System file verification never derives a percentage from sfc's output (not reliably parseable); it reports stage + elapsed. The bar blends the stage floor (started 10%, collecting 45%, uploading 85%) with elapsed-vs-estimate and is capped at 95% until the report actually lands — it can never show 100% early; a run that passes its estimate reads "Taking longer than usual — about N over the typical time". The estimate itself is labeled by source: the median of the tool's last 20 recorded run durations (recorded when a result lands, as result time minus delivered time) once at least 3 runs exist (history); before that, a per-tool rough guide set from the tools' real caps (fast checks ~30 s, Recent system errors ~1 min, Disk space ~2 min, System file verification ~10 min against its 30-minute cap).
- Copy customer update: each running card has a Copy customer update button that copies one ready-to-send sentence generated server-side from the live values — e.g. "We're running a drive health check on your computer now — it's about halfway and should take about 2 more minutes." — with distinct shapes for a queued run (starts at the device's next check-in), a run over its estimate (taking longer than usual), and a run with a real percentage (about halfway / nearly done).
- One shared name table: friendly tool names, stage phrases and the customer sentence all read from a single server-side table (
DIAG_FRIENDLY) that the diag endpoint ships to the panel, so the panel, the progress card and the sentence can never drift apart. Devices on older clients simply never send progress — their cards keep the previous queued/running text and reports land exactly as before.
Version 2.5.134
- Tech tools report live progress (with Server 2.5.169): the client posts stage updates to
/api/diag-progress while a check runs — started, collecting, uploading — plus the two honest extras: System file verification posts a still-running heartbeat every ~30 seconds carrying elapsed seconds (stage + elapsed only, never a percentage), and Disk space posts after each top-level folder completes with the real folders-done / folders-total counts (the one true percentage in the suite), and Recent system errors posts after each log channel. Posts are single-attempt and rate-capped (2 s minimum between posts unless the stage changed); a failed post is logged and forgotten — it can never fail, delay, or alter the diagnostic or its report, which still posts to /api/diag-result exactly as before. The checks themselves are unchanged: read-only, journaled on the device, never-invented values.
- Your Year in Protection (
GET /account/year-in-protection). A login-required report for one calendar year (default the current year; ?year= picks any other year the account has data in, with plain year links at the top). Every number is computed from the signed-in account's OWN rows — detections recorded against the account's installs during that year (detections.at, joined through installs.license_key), distinct threat names and per-device counts for that year, the device count, each install's lifetime_threats counter (labeled lifetime — the counters span all years and are never presented as per-year figures), and the account's created_at. No stat without an honest per-account source is shown: there are no stored per-year scan counts, so none appear. "Email me this report" (POST /api/account/year-report/email) sends the same summary plus a link back to the page through sendMail with kind year-report (journaled in the Email log); with SMTP unconfigured it fails visibly with a 400, like the support-reply flow. Linked from My Account (More tab) next to the protection summary. - Family member invites (Family tier only). New tables
family_invites (license_key, owner_account_id, invited email, token, status, 7-day expires_at, accepted_by/at) and family_members (UNIQUE license+member). From a Family license card on My Account the owner invites by email (POST /api/account/family/invite) — a non-family license is refused, as are the owner's own email and existing members; re-inviting revokes the previous pending invite. The invite email goes through sendMail with kind family-invite and links to GET /family/invite?token=…, which requires the invitee to sign in or register with the invited email (a signed-in account with a different email is refused 403 at POST /api/account/family/accept; expired tokens are refused and stamped expired; the owner cannot accept their own invite). Accepting creates the membership: the member's My Account gains a "Shared with you" section showing the license's status/renewal and its devices read-only (no action buttons are rendered for shared licenses). Enforcement is server-side, not cosmetic: delete-license, remove-device, transfer, switch-yearly and the invite-management endpoints all key on licenses.account_id, so a member hitting any of them gets 403. The owner manages members from the same card (GET /api/account/family/members, POST /api/account/family/revoke, POST /api/account/family/invite/revoke); revoking deletes the membership row and access ends immediately. Invite / accept / revoke are audit-logged. - Support screenshot attachments. The chat bubble (in the shared page chrome) and the
/contact form accept ONE optional image. Transports: multipart/form-data (what the browser forms send when a file is chosen — parsed by a small buffer-based parser in the new supportUpload middleware, since there is no multer dependency) or JSON image_base64. Validation is by magic bytes — PNG (89 50 4E 47…), JPEG (FF D8 FF), WebP (RIFF…WEBP) signatures only; the client filename/MIME are ignored, anything else is refused with a clear 400 (an EXE renamed .png never reaches disk), and the decoded cap is 5 MB (multipart bodies are additionally capped at 8 MB raw). Files land in DATA_DIR/support-uploads/ under a random hex name; support_messages gains attachment_file / attachment_type (both the chat insert and the contact mirror insert carry them). The admin Support inbox renders the image inline, served only by GET /api/admin/support-messages/:id/image behind adminAuth — there is deliberately no public/static URL for support uploads, and a logged-out fetch is refused. - PWA (site-wide).
GET /manifest.webmanifest is the primary manifest and GET /manifest.json now serves the same document (the 1.14.0 mobile manifest pointed at /mobile; both routes now serve the site manifest via one pwaManifest() helper): name/short_name AEAntivirus, start_url /, display standalone, background_color/theme_color #080b11 (the site chrome's --bg). Icons reference only files that exist in server/ today — /icon-192.png (verified 192×192), /icon-512.png (verified 512×512), /favicon-48.png. The public chrome (page()) now links the manifest, an apple-touch-icon, and a theme-color meta on every page, and registers the service worker. GET /sw.js keeps the 1.14.0 push handler and gains a real caching strategy: cache name aeav-site-<RELEASE_VERSION> (a release discards the old cache on activate), cache-first for static assets and the public help / tools / checker pages (/, /help, /tools, /link-checker, /scam-checker, /security-checkup, /threat-definitions), and network-only — never cached — for /account, /admin, /api/, /reseller, the auth pages, /vault, /widget/, cross-origin requests, and every non-GET request. The Help page explains adding the site to a phone's home screen as the user's own manual browser step (no iOS install prompt is promised anywhere). - Embeddable link-checker widget (
GET /widget/link-checker). A compact, chrome-less page (no site header/footer/nav/ticker/chat bubble — it does not use render()/page()) with the link-check form and result area, posting to the existing POST /api/link-check — the checker logic, the link-checker feature flag (off → a chrome-less unavailable notice, HTTP 200, and the POST refuses as before), and the per-IP rate limit all apply unchanged. ?brand=<reseller code or id> resolves against the resellers table and shows that reseller's name; an unknown brand falls back to AEAntivirus. Framing: the site previously had NO global framing protection (verified by grep) — this release adds it: every response now carries X-Frame-Options: SAMEORIGIN and Content-Security-Policy: frame-ancestors 'self' from a middleware at the top of server.js, except this widget, which gets frame-ancestors * and no X-Frame-Options (the handler re-asserts both), making it the only frameable page on the site. The /tools page shows the copy-paste <iframe> embed snippet under the link-checker entry, including the ?brand= note for resellers. - Homepage "New in" rotation keeps Server 2.5.134 / Server 2.5.133 / Client 2.5.115 visible with 2.5.132 folded into "See older updates". Help & FAQ gains entries for the Year in Protection report, Family plan sharing, support screenshots, adding the site to a home screen, and embedding the link checker (English; the Spanish help edition keeps its existing subset pattern — recent FAQ batches are EN-only). The admin guide documents screenshot handling in the Support inbox paragraph and the four customer-facing features in the System section. Regression coverage in
server/tests/server/v25134-batch7-test.js (84 checks: per-year aggregation against two seeded years + SMTP-captured year-report email + mail_log kind; the full invite → accept → read-only-view flow with member 403s on delete/remove/transfer/invite endpoints, revoke, expired token, and non-family refusal; PNG stored via both transports + byte-identical admin fetch, EXE-bytes and >5 MB refusals, logged-out fetch refusal; manifest/SW source + chrome-link assertions; widget chrome/brand/framing/flag assertions). Server-only release; the Windows client stays 2.5.115. Nothing for Austin to configure.
Version 2.5.168
- Marketing redo — antivirus + a technician, in one place (server-only; no client change): the homepage hero now leads with the technician half of the product — software protection backed by real technicians who can diagnose a device remotely, repair it in the shop, and recommend and install hardware upgrades (walk in, or mail it in). The "A real technician, built in" card is rewritten to name the real Tech tools checks, and Everything included gains a Technician services group: remote diagnostics, your service record, repairs, hardware recommendations, hardware upgrades installed, walk-in or mail-in. The Spanish homepage mirrors all of it (hero, card, group).
- New public page
/tech-services (English + Spanish, footer-linked, in the sitemap): how remote diagnostics work (the eleven read-only Tech tools checks, journaled on the device, records and fees under My Account → Billing → Diagnostics & repairs), repairs, hardware recommendations & upgrades, how mail-in repairs work step by step, and a walk-in card carrying the same shop facts the About page already publishes (Austin's Electronics, Harrells, NC). The page states only the shop's stated terms: hardware prices vary with the cost of the hardware and are always quoted before work; turnaround varies with the type of repair; mail-in customers pay shipping to the repair facility and back. No repair warranty, turnaround estimate, carrier, insurance, or data-handling promise was added anywhere — anything not stated is framed as confirmed with the customer before work begins. - Terms §3 gains a "Repairs, hardware work, and mail-in service" paragraph stating those same terms next to the service-fee rules (shop-set prices, draft invoices, never silently charged), so the marketing and the terms cannot drift apart.
- Help & FAQ gains a "Repairs & technician services" section in English and Spanish (remote diagnostics without bringing the PC in, how mail-in repairs work, what a diagnostic costs and how it is billed, how hardware is priced); the admin guide points staff at the page and §3 as the reference for customer conversations; and the My Account Diagnostics & repairs card intro now frames the list as the customer's service history with our technicians (data and account scoping unchanged).
Version 2.5.167
- Service fees for diagnostics & repairs (server-only; no client change): a fee ledger (
service_fees) for charging customers for Tech tools work. Diagnostic fee: an owner-set price in admin settings (diag_fee_cents, set under Customers → Service fees; default 0 — diagnostics are free until the owner sets a price; no price is ever invented). When a price is set and a diagnostic result lands (any of the eleven tools), an unbilled fee is auto-created for the device's account/install with the tool name and a link to the result. A run that errors or cannot run creates no fee — a result carrying an error, or an SFC verdict of failed/timeout, is skipped — and an identical unbilled fee for the same install+tool within 10 minutes is treated as the same run landing again (the client retries failed posts), never a second charge. - Repairs: on the install page, Record a repair — a description of the work, labor and parts in cents entered separately and summed (validated: whole cents, ≥ 0 each, total must be > 0 — a $0 repair is refused with a plain explanation), and an optional link to a prior diagnostic result on that install (a result from another device is refused). A repair is a record of work the shop performed — there is no client component and nothing is sent to the device.
- Billing out — drafts only: per-account Create Stripe invoice (Customers → Service fees) gathers the account's unbilled fees and creates a Stripe invoice with
auto_advance: false — a draft, one line item per fee (tool or repair description + date) — then marks the fees invoiced with the invoice id. It never finalizes, sends, or charges; the admin sends it from Stripe. If Stripe is not configured, or the account has no Stripe customer on file, the action refuses with a plain explanation and the fees stay as unbilled records. On any mid-flow Stripe failure nothing is marked (an orphan draft can be deleted in Stripe). Paid marking: the Stripe webhook's invoice.payment_succeeded branch now also updates service fees by invoice id outside the subscription check — one-off invoices have subscription: null, so the subscription-gated path could never see them; a payment marks the invoice's invoiced fees paid, audit-logged. - Customer visibility: My Account → Billing gains a Diagnostics & repairs card (
GET /api/account/service-fees, own account only): date, what it was (diagnostic tool or repair description), device, amount, and status in plain words — unbilled reads "Not yet invoiced", invoiced "Invoiced", paid "Paid", waived "Waived". Internal notes and who created a fee are never exposed. - Admin surfaces & roles: the install page shows the device's fee history plus the repair form; Customers gains a Service fees section (price setting, unbilled totals per account, full fee list, invoice + waive actions). All fee endpoints are
adminAuth and absent from the support allowlist — owner and manager, exactly like the install pages; the price setter is owner-only (requireOwner). Every create / price change / waive / invoice / paid transition is audit-logged. - Terms: §3 gains a service-fees paragraph in this release: prices are shop-set, diagnostics are uncharged until a price is set and never charged when a check cannot run, and service fees are invoiced as drafts the shop sends — never silently charged — matching the implementation.
- Tests:
v25167-fees-test.js drives real requests against the mock-Stripe seam (tests/server/stripe-mock-hook.mjs, extended in this release with invoices.create + invoiceItems.create — the established pattern; no live Stripe calls): fee auto-create at a set price, none at $0, none on error/failed results, 10-minute dedupe, waive flow, repair validation (negative + $0 refused, cross-device diagnostic link refused), draft-invoice creation with the exact line items and auto_advance false pinned from the mock's call log, the no-Stripe-customer refusal leaving fees unbilled, a mid-flow Stripe failure marking nothing, webhook paid transition for a subscription-less invoice, and customer-page scoping (an account sees only its own fees, with no internal fields).
Version 2.5.166
- Tech tools round 2 (with Client 2.5.133): six more read-only checks join the install page's Tech tools section — Startup items (Run/RunOnce registry values for machine + user including the WOW6432Node 32-bit view, user + common Startup folder items, and logon-trigger scheduled tasks read from the task files on disk so hidden tasks are included; every entry names its source, and the publisher is the target file's version-info CompanyName where resolvable within a bounded count), Hardware inventory (Win32_BaseBoard product/manufacturer/serial — OEM placeholder serials pass through verbatim, never sanitized into a fake; Win32_VideoController name/driver/adapter RAM; Win32_PhysicalMemory per slot — CPU stays in the Health snapshot, deliberately not duplicated), Windows Update (Win32_QuickFixEngineering installed KBs newest-first, capped at 60; last detect/download/install results from the Windows Update registry results keys; reboot-required state from the standard key probes; the pending count is reported as "not reported" — counting pending updates honestly requires a Windows Update Agent search, and this release deliberately does not run one or estimate a number), Recent system errors (Windows Event Log via wevtapi, System + Application, Critical + Error, last 7 days, grouped by provider + event ID with counts and the newest message — Windows' own formatted message where available, else the event's data values — top 25 groups, hard caps of 2,000 records scanned and 60 seconds), Disk space (bounded walk of the system drive — size per top-level folder plus its immediate children, reparse points never followed, caps of 250,000 entries and 120 seconds; a capped walk reports
partial: true and the panel labels the sizes lower bounds; folder-level only, no file names below the listed folders), and Security posture (firewall state per profile parsed from netsh's own all-profiles report — the same source family the isolation feature trusts; UAC EnableLUA; Secure Boot from the UEFISecureBootEnabled state key, an absent key reading "legacy BIOS or not reported"; BitLocker per volume from Win32_EncryptableVolume in its own WMI namespace — an unreadable namespace reports unknown, never "off"; and the antivirus products registered in root\SecurityCenter2, including whether AEAntivirus itself is registered). - Same rails as round 1: the six are new bare command types validated in all four places (server VALID_CMDS, remoteCommandIsValid, client cmd_parse_one, both client dispatchers) with the guard test extended to all eleven; results ride the same
POST /api/diag-result ingest (the result sanitizer's depth cap rises from 4 to 5 so disk-space child entries survive intact — still bounded by the key/array/string caps and the 32 KB serialized limit); the panel gains a readable renderer per tool (tables for startup/hardware/Windows Update/events, folder bars for disk space, labeled states for posture); access stays owner/manager via the install page — the support role has no installs access. - Privacy/Terms: Terms §4s and the Privacy Policy diagnostics bullet are extended in this release to name exactly what the six new reports contain (auto-start entries incl. command lines and publishers, board/GPU/memory details, installed-update lists, event-log message digests, folder sizes — folder names only — and the security posture summary), keeping the standing rules: read-only, journaled on the device, never-invented values.
- Two silent breakages fixed (same root cause): a redaction artifact baked into the source had corrupted two credential-adjacent strings. In the Windows client, the diagnostics report URL carried a literal placeholder instead of the device's install token — since client 2.5.132 every report a real PC tried to send would have been refused by the server (the server test suites simulate the device, so nothing caught it; fixed in Client 2.5.133 below, with a byte-level source guard). On the server, the email re-subscribe statement had lost its
CASE keyword the same way, so re-subscribing to the newsletter or outage alerts with an existing row threw a SQL error; fixed here, with the on → off → on cycle pinned by a real-request test.
Version 2.5.133
- Tech tools round 2 execution (with Server 2.5.166): the client runs the six new admin-requested checks and reports back over the same diagnostics channel. Fix included: the report URL now carries the device's real install token — a corrupted placeholder string in the 2.5.132 source meant reports from real PCs would have been refused by the server; the URL is pinned byte-for-byte by a source guard. Startup entries come from the registry (native + WOW64 views),
SHGetFolderPathA Startup folders, and the C:\Windows\System32\Tasks file tree (logon triggers and commands parsed from the task XML — the Task Scheduler COM service is deliberately not a dependency). The event-log digest loads wevtapi.dll dynamically (the same pattern used for srclient/dbghelp — no new link dependency), renders events to XML, and groups them locally with new unit-tested pure code (canned event records pin the parser, grouping, netsh firewall-state parsing, date keys and task-XML probes in tests/client/diag2-test). The posture check reuses the client's existing netsh capture path pattern; the BitLocker and SecurityCenter2 reads use their own WMI namespaces. All six checks are strictly read-only, every run is journaled on-device like round 1, and every value Windows does not report is labeled as such instead of being filled in.
- Broadcast announcement (Fleet → Broadcast announcement). A new admin composer: title (max 120 chars) + body (max 1000 chars) and an audience select — (a) all installs whose license is active or trial, or (b) one chosen license's installs (license picker lists every license with its status, install count, and whether it would actually receive anything). Sending stores a broadcast record in the new
announcements table (id, title, body, audience, license_key, target_count, created_by, created_at), writes an announce.send audit entry, shows the admin the enqueued count, and enqueues one announce remote command per target install via the existing JSON command queue (queueCommand, delivered by GET /api/commands). Excluded from both audiences: installs on expired, past_due, canceled, or deleted licenses, installs whose trial period has already ended, and installs with no matching license row — the bad states mirror licenseStateForClient. GET /api/admin/announcements backs the composer (past broadcasts + the picker); the send endpoint is admin-session + owner only and refuses empty titles/bodies, unknown audiences, and unknown licenses. - Wire format (new remote command type
announce). announce:<id>:<b64title>:<b64body> — the id is the decimal announcements row id (digits, no leading zero, 1..2147483647, same strict shape as hunt:/sandbox-file: ids); title and body travel as standard base64 of their UTF-8 bytes. Base64 was chosen because its alphabet contains no : (the three fields split unambiguously), no " and no control characters, so the payload survives the client's JSON-string command scan intact whatever the text contains — colons, quotes, and newlines in the body all round-trip byte-for-byte (proven by the test suite). Caps (title ≤120 decoded chars → ≤160 base64 chars; body ≤1000 → ≤1336) keep the whole command well under the client's CMD_TYPE_MAX (2048, unchanged — the type buffer holds the full command string). Four-place rule (server half): announce added to VALID_CMDS in server.js and to ARG_COMMANDS + remoteCommandIsValid in server/lib/upgrades-util.js (announcePayloadIsValid: strict id, canonical base64 round-trip, decoded-length and control-character checks; a bare announce with no payload is rejected even though the base is listed). A send whose text fits the character caps but not the wire caps (heavy multi-byte text) is refused with a clear error instead of queueing an undeliverable command. - Homepage "New in" rotation keeps Server 2.5.133 / Client 2.5.115 / Server 2.5.132 visible with 2.5.131 folded into "See older updates". Regression coverage in
server/tests/server/v25133-broadcast-test.js (58 checks: validation, both audiences, exclusions, refusals, record + audit, byte-identical round-trip through the real GET /api/commands fetch path). Nothing for Austin to configure.
Version 2.5.165
- Tech tools on the install page (with Client 2.5.132): every device's admin detail page gains a Tech tools section with five read-only diagnostic checks an admin can run against that one device: Hard drive check (each physical drive's model, serial, size and Windows status, the SMART failure-prediction state, and every volume's free space and dirty flag, with one plain health line per drive), Battery health (design capacity vs today's full-charge capacity as a wear percentage — desktops honestly report "no battery"), Connectivity check (from the device's side: gateway reachability, DNS resolution of aeantivirus.com, and a timed connection to this server), Health snapshot (CPU, memory, Windows build, uptime, volumes, and the five busiest processes from a 1-second CPU sample), and System file verification (Windows' own
sfc /verifyonly in the background — several minutes; the verdict is reported as no violations / corrupt files found / could not run, and verify-only means nothing is repaired). Results render as cards on the install page with their timestamps; the last 5 runs per tool are kept per device; a pending state shows while a check is queued or running, and results from before an offline device's last check-in are badged stale. - How it flows: the tools are five new remote commands (
diag-drive, diag-battery, diag-net, diag-health, diag-sfc) queued through the existing fleet command channel and validated in all four places a command type must exist (server command list, server command validator, client command parser, client dispatchers — a new guard test pins all four). The command acknowledgement channel only carries a short outcome string, so reports come back through a new device-authenticated ingest (POST /api/diag-result): results are deep-sanitized (plain data only, capped strings/arrays/depth), refused when the serialized report exceeds 32 KB, and stored per install (diag_results, newest 5 per tool). Tech tools follow the same access pattern as the rest of the install page: admins who can open an install (owner and manager roles) can run tools and read results; queueing rides the same endpoint as the other per-device commands. - Privacy/Terms: diagnostics are covered explicitly for the first time — a Privacy Policy bullet (§1) lists exactly what a diagnostic report contains, and Terms §4s states the rules: admin-initiated from the console only, read-only (nothing is repaired, deleted or changed on the device), every run is recorded in the device's own event journal, and results are visible in the admin console. Like the other fleet commands, a diagnostic runs when a registered device next checks in — the command channel is not license-gated (same as restart-device and reset-stats).
Version 2.5.132
- Tech tools execution (with Server 2.5.165): the client runs the five admin-requested diagnostic checks and reports the results back. Hard drive health comes from WMI (
Win32_DiskDrive + the SMART failure-prediction class — where a drive reports no SMART state the report says "not reported"; it is never invented), volume dirty flags from the filesystem itself (FSCTL_IS_VOLUME_DIRTY; "unknown" when the query isn't permitted). Battery wear compares the battery's designed capacity against its current full-charge capacity. The connectivity check resolves aeantivirus.com and times real TCP connections to the server and the gateway. The health snapshot samples process CPU over one second (labeled as such). System file verification runs sfc /verifyonly in the background and posts its verdict when it finishes — a 30-minute cap posts a timeout result instead of hanging forever. - On-device transparency: every diagnostic run writes a line to the client's own event journal naming the tool and that it came from the admin panel, so a run is visible on the device exactly like other fleet actions. All five checks are strictly read-only: nothing on the device is repaired, deleted, or reconfigured.
- Customer timeline (Customers → Accounts → account detail).
GET /api/admin/accounts/:id/timeline merges one chronological feed per account, newest first, capped at the latest 100 events (with total + showing in the response): account creation; license created / deleted events from the license rows; devices added (installs.created_at via the account's keys) and removed (revoked_installs rows for the account, with the removal reason); every mail_log row addressed to the account's email (kind + subject + sent/failed); admin notes added; and audit.log entries attributable to the account — by actor (customer:<email>), by a bounded account <id> reference (refunds log "for account <id>", notes log "account <id> note #…"), or by containing one of the account's license keys (transfers, trial conversions, subscription cancels). Broadcasts reach the feed through their per-recipient mail_log rows (kind broadcast) — there is no separate per-account broadcast record, so none is fabricated. The detail view renders every field through esc() (subjects and notes are attacker-influenced text), in its own independent loader. - Config & connections viewer (System → Configuration).
GET /api/admin/config-status is presence-only: SMTP_HOST / SMTP_PORT / SMTP_USER / SMTP_PASS / SMTP_FROM / SMTP_TO (+ overall mailConfigured()), STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET, STRIPE_PUBLISHABLE_KEY, and every Stripe price-ID env var the code reads — the fallback + all seven tier prices, the six yearly tier prices, and the eight add-on prices (vault, leak alerts, identity, VPN, AI assist, DDoS, managed, storage, disk cleanup) — each by env NAME with Set / Not set and a one-line "what breaks without it" hint, grouped Email / Payments / Integrations / Server. Integrations adds OPENAI_API_KEY (+ HIBP_API_KEY); Server adds PUBLIC_URL, the DATA_DIR path with a real writability check (fs.accessSync W_OK), and session-secret presence (the secret is a server-side file under DATA_DIR, never an env var, and its value is never returned). No env value is ever included in the response. - Trial abuse watch (Customers → Trial abuse watch). What provisioning recorded before this release:
trial_registrations rows with IP + device fingerprint + time only — no email, no key — and trialRateOk() pruned rows older than 24h on every check, so no usable history survives from before 2.5.132. This release instruments provisioning going forward: registrations now also store the account email and the trial key minted (both paths — the /download//account ensureTrialKey() mint, which records by IP, and the installer /api/register mint, which records IP + fingerprint), and the 24h prune is removed (the rate counts already window-filter, so retention costs the limiter nothing). GET /api/admin/trial-watch groups signups by IP and by fingerprint, repeats only (count > 1, sorted by count), with emails, keys, and first/last times, and states the tracking start honestly. A new trial_blocklist table (kind ip|fp, value, created_at, note) is consulted by both provisioning paths before minting: a blocked IP/fingerprint is refused with the exact same generic message as a rate-limited trial — no new leak about why. Block/unblock are owner-only and audit-logged (trial.block / trial.unblock). - Export center (Customers → Export center). Owner-only, every export audit-logged (
export.* with the row count): Customers CSV (email, name, created_at, health rating, license count), Licenses CSV (key masked to + last 4 only, tier, status, platform, created/expires, account email), Installs CSV (hostname, platform, masked license, first/last seen, lifetime threats), Detections CSV (time, hostname, threat, verdict/status, path), and Invoices CSV via Stripe (last 12 months: date, customer email, amount, status, invoice id — gathered per Stripe customer id from license/add-on rows, so it runs against the real API and the stubbed-Stripe test harness alike; with Stripe unconfigured it returns a clean 400, not a broken file). Every cell in every export passes a formula-injection guard on top of CSV quoting: a cell whose first non-space character is =, +, -, or @ gets a leading apostrophe. A full license key never appears in any export. Homepage "New in" rotation keeps 2.5.132/2.5.131/2.5.130 visible with 2.5.129 folded into "See older updates". Server-only release — client stays 2.5.114-WC, no installer or update-manifest changes, nothing for Austin to configure.
Version 2.5.164
- Memory soak baseline (admin, System → Server health): a new soak utility (
server/tests/scale/soak.js) seeds the bench dataset — 5,000 installs, 50,000 detections — boots the real server, and loops every admin endpoint as owner while sampling process memory. Its report ships with this release and is shown read-only next to the live server-memory card, labeled with its date, dataset and "sandbox baseline — compare with the live numbers above". There is deliberately no "run soak" button in the panel: a soak is too heavy for the production process. Sandbox result for this build: RSS 127.3 MB at start, 208.6 MB peak, 208.6 MB at end across 177 endpoints × 3 rounds; heap used 28.5 MB → 54.8 MB. The full per-endpoint table ships in server/soak-report.md. - Bug fix the soak caught: the scam-call blocklist's admin endpoints (add/list/remove numbers) returned an error page to a signed-in admin on every call — they had been checking admin rights through an outdated call pattern since the sign-in system was reworked. Anonymous visitors were correctly refused, which is how it stayed hidden. All three endpoints now work, with a regression test.
- Public status page — verified, unchanged: the page at
/status was audited for this release and already does what it should: live self-measured checks, owner-posted incidents with their investigation timeline, uptime computed from the incident records themselves (labeled with how far back the records reach), and an outage-alert signup. A new end-to-end suite pins that behavior: no incidents → "No active incidents.", posting → reflected on the next request, resolving → moves to history.
Version 2.5.163
Batch 21 — a server-only release (the Windows client stays 2.5.131, Android 1.31.0, Linux 1.25.0, extension 1.2.0): an owner "waiting on you" card, support-inbox aging, the Threat Guide in Spanish, a golden-path end-to-end suite, and a repository cleanup.
- "Waiting on you" card (item 1). Dashboard → Overview gains a card computed fresh on every load from live state (new endpoint
GET /api/admin/waiting-on-you, admin session only — nothing hardcoded). Open gates, each with a plain-words state, what it unlocks and where to fix it: SMTP not configured; Stripe secret key missing; Stripe webhook secret missing; STRIPE_DDOS_PRICE_ID unset; off-site backups not configured; zero approved customer reviews (homepage reviews section hidden); paid plan/platform mismatches pending; reseller custom domains unverified; verified reseller domains whose TLS still needs adding in the Render dashboard (stated honestly — the server cannot see Render's list, so the row persists while any verified domain exists); trial-abuse blocklist non-empty. Closed gates collapse into an "Already handled" line. The mismatch check reuses one shared collectPaidMismatches() helper extracted from the license-mismatches endpoint (its behavior is test-pinned unchanged). - Support inbox aging (item 5).
GET /api/admin/support-messages rows now carry age_hours (computed at request time) and the response carries a filter-independent waiting_count; new sort=oldest (default stays newest-first, i.e. arrival order) and waiting=1 (open tickets older than 48h) parameters. The inbox UI shows a humanized age per ticket, an amber "Waiting Xd Yh" badge on open tickets past 48h, a "Waiting >48h" filter chip with the live count, and a sort control. - Threat Guide in Spanish (item 6).
/threat-definitions now honors the site's language selection (?lang=es or the aelang cookie) and renders the complete guide in Spanish: all 84 threat entries across the 5 groups, plus every framing section (attack chain, signs, what-to-do, myths, glossary, alert decoder, red flags, damage control, and the 10-question quiz with a Spanish scoring script). Data lives in server/views/definitions-es.js, mirroring GUIDE_GROUPS slug-for-slug; the English render is byte-identical to before. Fallback rule: per-group completeness — a group renders in Spanish only when its mirror has the same slugs in the same order and count, otherwise the whole group renders in English with a visible "Traducción en curso" badge and a notice card; an entry is never half-and-half. Feature names (Ransomware Rollback, Boot Guard, Smart Scan…) and terms like ransomware/phishing/malware stay in English, consistent with the site's existing Spanish pages. - Golden-path E2E suite (item 2).
server/tests/server/v25163-golden-path-test.js walks one scripted journey against a scratch server — register → trial mint follows platform (Android → Mobile plan, Windows → generic trial) → device register → heartbeat (paid add-on flows; license reads trial) → detection WITH a signal snapshot → device false-positive verdict → admin ML eval counts it (verdict honestly NOT EVALUABLE at one sample; the snapshot-only count proves the snapshot is what makes the sample usable) → exclusion syncs back via GET /api/fp-exclusions in exactly {path, threat} shape → expired heartbeat (add-ons emptied, grace anchored to plan end + 72h) → update check returns the curated manifest (2.5.131, never the server version) → rollback arm ingested once, deduped, and read back. Every assertion names its seam. - Repository hygiene (item 3). The stale root
lib/threat-util.js and lib/upgrades-util.js duplicates are removed (verified: nothing loads them — the live copies are server/lib/, 17,409 B and 25,321 B vs the stale 4,095 B / 14,754 B root copies; the repo root drops from 19 to 18 top-level entries). Every local-only test file was dispositioned: the drifted v1389/v1581 server suites (v1389 needed one comment-only fix — two em dashes in src/tray.c comments replaced with -- so its ASCII pin passes), four local-only client suites (honeypot, ioc, netbehavior, rstream), the v2400 cloud/heur suites, the Linux v1200 suite plus remotestream.c/.h (the repo's own Linux compile lines already linked remotestream.c — the pushed tree could not build the daemon without them), and 29 suite-generated *_block.inc fixtures are now tracked; the generated detection-dataset samples are NOT tracked (the tracked generator.js recreates them; EICAR strings stay out of the repo) and are gitignored instead. - Linux test legs repaired (item 4). The daemon compile lines in
linux/v1240-test.sh and linux/v1200-test.sh were missing sigdelta.c (the AESIGDELTA engine added in Linux 1.25.0), so the compile leg failed on undefined aesig_* references — v1250's line already had it. Both lines fixed; the batch13r4 Linux leg (which delegates to v1240) now runs for real: 7/7 unit checks, daemon compiles and prints 1.25.0, 6/6 guards. The old "headers missing in this sandbox" skip no longer applies (libcurl/OpenSSL headers are present); v1240's skip branch now prints a loud reasoned SKIP: line if headers ever go missing again. - Tests:
server/tests/server/v25163-batch21r1-test.js (items 1+5: every gate flipped individually from both sides — bare env, seeded rows, a full-env second boot — plus the inbox filter/order/age pins, 49 checks), server/tests/server/v25163-golden-path-test.js (item 2, 40 checks), server/tests/server/v25163-batch21r2-test.js (item 6: ES render over HTTP, slug-for-slug data completeness, stubbed-module fallback, inline-script parse on EN+ES pages, 46 checks).
Version 2.5.130
Batch 19, release B — the Windows client half of the batch: crash reports (item 2, client), the detection signals snapshot (item 1, client), and the Disk Tools restore-points panel (item 6). Windows client only; the server stays 2.5.161, Android stays 1.29.0, Linux stays 1.25.0.
- Crash reports (item 2, client half). The 1.8.0 crash logger only wrote the exception to the local daily log — the "awaiting a log tail from Austin" problem. 2.5.130 extends the existing unhandled-exception filter (no second system): after logging, it also writes a minidump (dbghelp
MiniDumpWriteDump, loaded dynamically) and a pending-report metadata file (client version, exception code/address, faulting module, the in-flight operation, timestamp) into %PROGRAMDATA%\AEAntivirus\crash\. A last-actions ring buffer (crash\actions.log, newest 200 lines / 32 KB, timestamps + action names only — no file contents, no typed text) is fed by the existing crash_op breadcrumb (on change, throttled) and by tray page opens. On the next successful service start, pending reports upload to the 2.5.161 endpoints — metadata + actions tail first, then the dump — and the local copies are deleted only after the server acknowledges; at most 3 reports per start, and a report that keeps failing is dropped after 5 attempts. The upload plan (oversize dump → metadata only, since the server caps dumps at 5 MB) and the ring buffer are unit-tested (server/tests/client/batch19r2-test.*); honest limit: a real crash's end-to-end proof is Austin's machine, not this VM — the capture path is guarded/fail-silent by construction and source-pinned, but no Windows box here ever crashed with it. - Detection signals snapshot (item 1, client half). When the client reports an
AI.Heuristic detection it already scored, the report now carries a signals object — the per-signal scores from the same fired-signal set and active weights the score_detail breakdown string shows (aiscore_last_signals_json: positive-weight fired signals only, strongest first, ≤48 entries, bounded JSON), so the snapshot can never disagree with the breakdown. Server 2.5.161 stores it on the detection row and the ML corpus counts it (round trip proven in server/tests/server/v25161-batch19r2-clientrt-test.js with the exact client payload shape). - Restore Points panel (item 6). Disk Tools gains a Restore Points (Created by AEAntivirus) panel listing the pre-remediation restore points from the client's own event journal (each creation is journaled with its Windows sequence number) — description ("AEAntivirus before threat removal") + timestamp; points created by Windows or other apps are never listed. Delete uses the supported API:
SRRemoveRestorePoint (srclient.dll, dynamically loaded — the removal counterpart of the SRSetRestorePoint used to create them), keyed by the journaled sequence number; the deletion is journaled, so the panel stops listing it. No shadow-copy enumeration or guess-matching anywhere. The page's sections were tightened to fit the panel (same tools, slightly denser layout — Austin's on-screen check pending, as with every visual change). - Tests:
server/tests/client/batch19r2-test.sh (ring buffer trim/tail/sanitization, metadata round trip, upload-plan matrix, signals-JSON ordering/caps/weights — 228 checks incl. source guards) and the round-trip suite above. Build: x64 then x86 sequential; both EXEs verified from disk (embedded 2.5.130, crash + signals markers present). Manifest curated: version/notes/exe hashes from this build; sigs/certs/trusted-signers channel fields from the GitHub manifest (the local tree's manifest still carries the stale stub sigs hash — not used).
- Customer health ratings for direct customers (Customers → Accounts). Every direct (non-reseller) customer now carries a healthy / attention / critical badge plus a rating filter. The definition is NOT new: it is the reseller portal's customer-health logic, extracted into one shared
computeCustomerHealth(accountId, resellerId) helper in server.js that both surfaces now call — critical = any license in expired/cancelled status, or devices exist but none has heartbeated in 24h; attention = any stale device (no heartbeat for 7+ days), paused protection, detections in the last 30 days, or an active license expiring within 7 days; otherwise healthy. GET /api/admin/accounts returns health + health_detail per direct account and accepts ?health=healthy|attention|critical to narrow the list server-side. - Bulk email composer (Customers → Bulk email). Subject + body, segment select —
paid (all active paid), trials, expiring (active licenses ending within 14 days), custom (Custom plan owners) — computed as distinct account emails from license rows. GET /api/admin/broadcast/preview?segment= returns the exact recipient count (count only, no send); POST /api/admin/broadcast/test sends only to an admin-entered test address; POST /api/admin/broadcast/send requires confirm_count to equal the current recipient count (the UI confirm shows that exact count), sends per recipient through the mailer with kind broadcast (one mail_log row each), returns sent/failed counts, and writes an audit entry. When SMTP is not configured the composer is visibly blocked with the reason and the send/test endpoints refuse with the same reason — no silent failure. - Stripe refund tool (Customers → Accounts → Billing).
GET /api/admin/accounts/:id/charges lists the customer's recent Stripe charges (brief in-memory cache) and POST /api/admin/accounts/:id/refund (owner-only) issues a full-amount refund for one paid charge, returning Stripe's refund status and writing an audit entry with the charge id and amount. The UI double-confirms and states plainly: "A refund does NOT cancel the subscription — cancel it separately if needed." Already-refunded charges are refused gracefully; with no Stripe key or no charges the view shows a clean message, not an error. - Webhook event log (Customers → Stripe events). The Stripe webhook handler is wrapped so every verified event is recorded in the new
stripe_events table (event id UNIQUE, type, Stripe created time, received time, handled/failed + error text, related subscription/customer/invoice id, payload). A Stripe retry of the same event id increments a deliveries counter and refreshes received_at — never a duplicate row. Logging is best-effort and never changes webhook behavior: the handler responds (and fails) exactly as before. The admin viewer searches by type / related object, filters by status, and expands a row to the payload JSON. stripe_events is registered in the Data cleanup prunable registry. - Feature flags (System → Feature flags). New
feature_flags table (key, enabled, updated_at, updated_by), seeded all ON: chat, link-checker, scam-checker, checkup, reviews, referrals, and deals (pre-created for the next release; no surface yet). Owner-only toggles, audit-logged, read through a ~3-second cache. When a feature is OFF: its public page renders a friendly "temporarily unavailable" notice (HTTP 200, not a bare 404), its footer links are hidden, and its POST endpoints refuse with a clear JSON error; chat off removes the bubble markup from served pages entirely; reviews off omits the homepage reviews section; referrals off hides the /account referral card and disables referral discounts at checkout — non-referral checkout works identically. Server-only release; the Windows client stays 2.5.114.
Version 2.5.131
Batch 20, release B — the Windows client half of the batch: battery-aware scheduled scans (item 6). Windows client only; the server stays 2.5.162, Android stays 1.30.0, Linux stays 1.25.0.
- Battery-aware scheduled scans (item 6). Until now the scheduler's own header said it plainly: "the scan runs regardless of AC/battery state." 2.5.131 adds a battery gate to the scheduled-scan module (
src/schedscan.c), consulted by BOTH scheduled paths — the local daily/weekly schedule and the fleet policy schedule. When a scheduled scan comes due and the machine is on battery below the threshold (config schedscan_battery_pct, default 25), the scan defers: the deferral is journaled exactly once per episode (scheduled scan deferred — on battery at N% in the event journal), power is re-checked on the normal 60-second poll, and the scan starts as soon as AC returns or the battery recovers to the threshold. A deferral is capped at 24 hours — at the cap the scan runs regardless, journaled as a cap run — and the deferral start persists in the stats file, so the cap is honest across restarts. For the fleet schedule, a deferred slot is remembered (ssched_deferred_slot) and fired when power allows even after its 15-minute window has closed (dropped, not fired, if the policy is disabled or the license lapses meanwhile). A user-started scan is never deferred — user scans do not pass through the scheduler at all, and the pure decision function takes the scan origin as an input so the rule is unit-tested, not just wired. The decision (schedscan_battery_defer, in the existing testable block) defers only when origin = scheduled AND power is KNOWN to be battery AND the percentage is KNOWN (unknown power / no battery / unknown percentage all run the scan) AND percent is strictly below the threshold AND the deferral age is under the cap; threshold 0 disables the gate. Wiring reads power via GetSystemPowerStatus (ACLineStatus, BatteryLifePercent, BatteryFlag 128 = no battery). Overlap note: the Scan page's older "Battery-aware" display (smartscan.c, a fixed 20% indicator whose should_pause verdict nothing consumed) is a status display only — 2.5.131 is the first actual deferral behavior, and it lives in the scheduler where scheduled scans live. - Tests:
server/tests/client/batch20r2-test.sh compiles the testable block and runs the full decide matrix (origin × power state × battery % × threshold × defer age × cap, incl. recovery mid-deferral) plus source guards (GetSystemPowerStatus wiring, the exact journal line, the gate on both scheduled paths, deferred-slot persistence, version + What's New). The existing polsched suite (fleet slot rules) and the guide-format test re-run green against the changed block. Build: x64 then x86 sequential; both EXEs verified from disk (embedded 2.5.131, journal string + What's New present, manifest hashes match the files). Manifest curated: version/notes/exe hashes from this build; sigs/certs/trusted-signers channel fields from the GitHub manifest.
- Family Safety Center (/family). A new parent-focused page (footer → Products) that tells one story across the four family features that already ship in the Windows app, with setup steps for each: Parental Controls (category blocking — adult, gambling, social, gaming, streaming, dating — enforced at the DNS layer, per-day internet windows, tamper-PIN-gated settings, and a local on-device activity log), the DNS content filter (malicious/phishing blocklists plus ad & tracker blocking for the whole PC), Child-safe search (the paid add-on that forces SafeSearch on major search engines via DNS), and Focus Mode (the app's Focus & Safety page — session site blocks via the hosts file plus distracting-app closes), plus per-app time limits. Copy names only shipped behavior, links the User Guide's Family controls section, and states plainly that filtering decisions happen on the family PC — browsing history is never sent to the server.
- Business quote calculator (/business). A device-count control (stepper + number input, 1–100) with a monthly/yearly toggle. Every count's per-device rate, monthly total, and yearly total are computed server-side from the same sources checkout bills against —
CUSTOM_PER_DEVICE_CENTS, CUSTOM_VOLUME_TIERS, customEffectivePerDeviceCents(), customVolumePct() — and embedded as a lookup table the calculator script only reads, so the page can never quote a stale or re-derived price. Yearly, honestly: the Custom plan has no yearly Stripe price (checkout always bills it monthly per device; yearly prices exist only for fixed tiers, as the 2.5.120 post-purchase switch), so the yearly figure is the annual cost — 12 × the monthly total at the same volume-discounted rate — and the calculator, the subscribe preset note, and the checkout metadata (requested_interval, recorded for reference and changing no pricing) all say exactly that. The CTA links to /subscribe?tier=custom&devices=N&interval=monthly|yearly, and /subscribe now honors the preset: the pay card renders open with the tier pre-selected, the Custom device count set, the volume-aware total pre-computed server-side, and a plain note when the yearly view was chosen. - Trust Center (/trust). A new factual page (footer → Legal): what data is collected (account — email, name, scrypt-hashed password; license — key, status, Stripe identifiers; device/install — hostname, versions, hashed device fingerprint, module states, scan counts; detections — threat name, file path, time), where it is stored (the server database on Render; payments with Stripe; email via the configured mail provider), retention from the real code (automatic retention purge off by default, 90-day default windows when enabled, Extended-history accounts at 365 days, email log pruned after 90 days, audit log keeping its newest 5,000 entries, Data cleanup prunes of 30/90/180 days for operational stores only — licenses, accounts, devices, and payment data are never prunable), and the security practices actually in place (scrypt hashing + breached-password refusal, TOTP 2FA with hashed recovery codes, passkeys, audit logging, configuration shown as set/not-set with API tokens stored as SHA-256 hashes and shown once, and the one-click backup download's standing exclusions — session secret, 2FA material, vault contents, .env files). No certifications are claimed and no compliance badges are asserted — only "we do X" statements. Links to /privacy and /terms.
- Deals page + promo public flag (/deals). The Stripe promo manager (Customers → Promo codes) gains a per-code Public toggle and public description, persisted in a new local table
promo_public keyed by Stripe promotion code id (Stripe has no such flag; Stripe remains the source of truth for the code, discount, active state, and expiry) — settable at creation or from the list, owner-only and audit-logged. /deals lists a code only when it is flagged public AND Stripe-active AND unexpired AND under its redemption cap, showing the code, the discount (percent or fixed amount, from the Stripe coupon), the description, and the expiry; with no qualifying codes it renders a clean "No public offers right now" state (still 200, also when Stripe is unconfigured or unreachable). The page honors the pre-created deals feature flag: off → the friendly unavailable notice, and the footer's Deals link hides (same pattern as the other flagged links). Footer also gains Family Safety Center (Products) and Trust Center (Legal). - Outage alert signup (/status). The status page gains an email signup card (honeypot + per-IP hourly rate limit, same pattern as the Scam Alert newsletter) backed by a new table
status_alert_subs (email UNIQUE, token, created_at, unsubscribed_at) with a friendly, idempotent GET /status-alerts/unsubscribe?token= page. The existing incident flow now also emails subscribers: posting an incident from System → Service status sends "Service incident: <title>" and resolving an active one sends "Resolved: <title>", both through sendMail with kind status-alert (so every attempt is in the Email log), each carrying the subscriber's own unsubscribe link. Duplicate signups never duplicate a row (re-signup after unsubscribing revives the row with a fresh token), unsubscribed addresses receive neither email, re-resolving an already-resolved incident sends nothing, and a mail failure never fails the incident post. Help/FAQ gains entries for family safety, the Trust Center, deals, outage alerts, and the quote calculator; the homepage "New in" rotation keeps 2.5.131/2.5.130/2.5.129 visible with 2.5.128 folded into "See older updates". Server-only release — client stays 2.5.114-WC, no installer or update-manifest changes, nothing for Austin to configure.
Version 2.5.162
Batch 20, release A — the server half of the batch: backup restore (item 1), fleet tags completed (item 3), the reseller provisioning API completed (item 4), crash-signature alerts (item 5), database maintenance (item 7), and the server half of extension false-alarm reports (item 8). Server-only release; the Windows client stays 2.5.130, Android stays 1.30.0, Linux stays 1.25.0, the extension stays 1.1.0 (its half of item 8 ships as extension 1.2.0).
- Backup restore (item 1). Until now a backup could be verified (the 2.5.160 drill) but never restored from the panel.
GET /api/admin/backups/restore-info (owner) lists every local backup with its drill-verification status (a verification only counts while it is at least as new as the backup file). POST /api/admin/backups/restore-local (owner, typed RESTORE confirmation) requires a passing verification of THAT backup — if it has none it runs the drill inline first, and a failing backup is refused with nothing staged. It then stages the restore: the backup's database is extracted beside the live one, a safety snapshot of the live database is taken first (a normal backup in the same list, itself restorable), a boot flag is written, and the process exits for the platform restart. At the next boot, before the database is opened, the staged file is integrity-checked again and swapped in; the replaced database is kept as aeantivirus.db.replaced-<stamp>. Failure behavior: any failure at staging or at boot discards the staged file and the server boots on the current database, untouched — a staged restore can never lose the live data. Only the database file is restored (backups are whole-data-dir tarballs, but settings/policy files are left alone). Both ends are audit-logged (backup.restore.staged, backup.restore.applied). If the host does not restart the process (Render does), the staged restore applies on the next manual boot — the admin copy says exactly that. The older 1.17.0 raw-upload restore endpoint is unchanged. - Fleet tags completed (item 3). Tag storage and the per-install editor pre-existed (1.4.6); the fleet-level half was missing. The installs API now takes an exact
?tag= filter (the fleet's old substring box filtered only the loaded page), POST /api/admin/installs/bulk-tags (owner) adds/removes/replaces tags on an explicit selection (≤500) or on every install carrying a tag, and bulk fleet actions accept a tag target: the tag resolves to all installs carrying it at send time with no 50-cap (explicit selections keep the cap), the fleet bar's new by-tag buttons confirm with the live match count first, and the audit entry names the tag. - Reseller provisioning API completed (item 4 — audit first, per the steering note). Audit verdict: the key system and the provisioning path already existed (1.17.0: scoped tokens —
read/commands — shown once, SHA-256 at rest, revocable, last-used; POST /api/reseller/licenses/issue already authorized tokens through the same resellerTokenOrSession('commands') middleware and the same key_quota check as the portal). Building a second key system — or a redundant provision scope over an endpoint commands already reaches — would have fragmented the scope model, so none was built. The genuine gaps closed: a per-key rate cap (300 calls / 5 minutes per key → 429), an audit-log entry for every token-authenticated call carrying the key id and name, and the missing single-license status check GET /api/reseller/licenses/status?key= (read scope; another reseller's key reads as not found). The reseller portal's token card now documents the API. Scope isolation is pinned by tests (read token cannot mint, another reseller's licenses are invisible, a revoked token is refused, token and portal mints count against the same quota). - Crash-signature alerts (item 5). Crash reports now store a signature at ingest — exception code + faulting module + client version family (e.g.
0xC0000005|aeantivirus.exe|2.5.x) — computed from the metadata the 2.5.130 client already sends. When one signature reaches ≥3 reports from ≥2 distinct devices within 7 days, exactly one alert is filed into the Support inbox (source crash-alert); further reports of the same signature update that alert's counts instead of creating rows, and archiving the alert lets a later threshold hit file a fresh one. Pre-2.5.162 crash rows carry no signature and do not count. - Database maintenance (item 7). New System card (next to Retention): live database + WAL sizes, the last vacuum (before → after, duration, reason), an owner-only VACUUM now, and a toggle to auto-vacuum after a retention purge that deleted more than 10,000 rows (the purge response then carries the vacuum result). VACUUM + ANALYZE rebuild the database file so purge-freed space returns to the OS; it holds an exclusive write lock for its whole run (writes pause, reads continue) — the measured duration on this VM's seeded test database is reported in the release notes to the owner; on a production-size database it should be run deliberately, which is why the automatic path only fires after a >10k-row purge. No chunked/incremental vacuum exists in SQLite; the honest trade-off is stated in the card copy.
- Domain false-alarm reports (item 8, server half). New public endpoint
POST /api/domain-report — no device auth (the extension's block page has none), 10 reports/hour/IP. The domain is validated and normalized (lowercased, trailing dot stripped, scheme accepted only from a bare URL — any URL carrying credentials is refused outright, as are paths, ports and whitespace), and stored into the same verdict_reports queue the verdict tools use (tool domain, source marked extension), triaged in the existing admin Verdict reports view. The extension half (the block-page button) ships as extension 1.2.0. - Tests:
server/tests/server/v25162-batch20r1-test.js (restore round trip across two boots incl. safety snapshot + corrupt-backup refusal, tag CRUD/filter/bulk-by-tag set equality, reseller scope isolation + quota parity + revocation + rate cap + audit key ids, crash-alert thresholds + dedupe + refiling after archive, VACUUM before/after + auto-vacuum after a >10k purge, domain-report accept/normalize/refuse/rate-cap/triage, inline-script regression on all pages).
Version 2.5.161
Batch 19, release A — the server half of the batch: detection signal snapshots for the ML corpus (item 1, server), crash-report ingest + admin view (item 2, server), the live Server memory card (item 3), per-class retention windows (item 4), reseller-branded customer email (item 8, audit first), and the printable emergency one-pager (item 9). Server-only release; the Windows client stays 2.5.129, Android stays 1.29.0, Linux stays 1.25.0.
- Detection signal snapshots (item 1, server half). Verified before-state: a recorded detection verdict carried only the fired tag names in
reasons plus a compact score_detail string — no per-signal score snapshot was stored anywhere, which is why the ML corpus could count verdicts but never reconstruct their features. 2.5.161 adds a signals column to detections: POST /api/detection accepts an optional signals object, sanitized server-side (plain object only; keys lowercased [a-z0-9-]{1,40}; values finite numbers clamped 0–10000 and rounded; ≤64 entries; ≤4096 bytes serialized; anything malformed degrades to *no snapshot*, never an error). The corpus evaluator (live library + standalone harness, kept in lockstep) now unions snapshot tag names into a verdict sample's feature set and reports how many verdict samples carry a snapshot and how many are usable only because of it; the admin ML panel shows the same counts. Backfill is impossible and the panel says so: detections recorded before client 2.5.130 / server 2.5.161 have no snapshot — the scores were never stored, only the fired tag names in reasons. - Crash-report ingest + admin view (item 2, server half). New device-authenticated endpoints (own body parsers ahead of the app-wide JSON parser, the quarantine-upload pattern):
POST /api/crash-report stores metadata + a capped last-actions tail (≤40 actions, ≤200 chars each) and returns a dump URL; POST /api/crash-report/dump?id= accepts the raw minidump (≤5 MB, stored on disk under DATA_DIR/crash-dumps/, never base64 in SQLite; empty → 400, oversize → 413, another device's report → 404). Per-install cap: 5 reports/hour (6th → 429), so a crash loop cannot flood storage. System → Crash reports lists device/version/when/exception with per-report detail (full actions tail) and a dump download. The client half (capture + upload) ships in 2.5.130. - Server memory (live) (item 3). A lightweight ring keeps the last 300 request durations (route keys collapse ids so cardinality stays bounded; only method + collapsed path + milliseconds are stored — the instrumentation cannot become the memory problem it measures).
GET /api/admin/server-memory returns RSS / heap used / heap total / external from process.memoryUsage(), uptime, Node version, and the slowest endpoints (count, average, worst). The System tab's Server health section gains a Server memory (live) card labeled as this server process's own numbers. - Per-class retention windows (item 4). The retention system is rebuilt around classes: detections, device events (
security_events), mail log, webhook deliveries, feed pull stats, heartbeats, and the audit log file — each with its own window in days, 0/blank = keep forever (the default for every class; nothing is deleted until the owner sets a window). The purge runs on the existing jobs runner plus the System tab's Purge now button, deletes in 1,000-row chunks, and logs per-class counts. The System → Retention card now shows each class's live row count, oldest row, and how many rows the saved window would remove — the effect is visible before anything is enabled. False-positive exclusions and licenses are never purge classes (pinned by test). This also repairs the tab itself: it previously read a response shape the API never sent and its Save button called a function that did not exist. - Reseller-branded customer email (item 8 — audit first). Audit verdict: the gap was real. Web pages brand per account (
branding table) and per reseller domain, but no email send site consulted either: every template default says AEAntivirus and the From was the bare SMTP_FROM address. 2.5.161 makes sendMail brand-aware for customer mail: sites opt in per recipient, the brand resolves as the account's own white-label product name first, else its reseller's name; with a brand, the From gains the brand as its display name and the product name in the subject/body/sign-off follows the brand (links and aeantivirus.com addresses stay as-is so they keep working). Wired at the customer send sites: verification, password reset + password changed/reset notices, expiry/trial/dunning reminders, win-back, gift licenses, detection notices, threat alerts, family digest/summary/invite, fleet report, customer digest, year report, support replies, broadcasts (per recipient), breach alerts, and the site/SSL monitor alerts. Unattributed accounts resolve to no brand and their mail is byte-identical to before; admin/internal mail (SMTP_TO alerts, subscriber lists, phishing-training simulations) is untouched. Proven in the release suite via the mail journal: a reseller-attributed account's reset email is journaled with the brand in the subject, a direct account's with AEAntivirus. - Printable emergency one-pager (item 9). New public page
/emergency — "I'm infected — do this now": the Threat Guide's two damage-control lists ("What to do if you think you're infected", "If you already clicked or paid") reproduced verbatim on one print-first page (print CSS hides the site header/footer/nav and forces a clean light sheet that fits one page), AEAntivirus branding, a Print button, and the short URL aeantivirus.com/threat-definitions back to the full guide. Linked from the Threat Guide and from a new Help & FAQ entry. No QR code: the server bundles no QR library, so the sheet carries the short URL as plain text rather than adding a dependency for one image. - Tests:
server/tests/server/v25161-batch19r1-test.js (signals round trip + corpus counting incl. a snapshot-only sample, malformed/oversize degradation, crash round trip incl. oversize 413 + 6th-in-hour 429 + cross-device 404, server-memory shape, retention per class with FP exclusions/licenses untouched, mail-brand journal proof + From-header source pin, /emergency smoke incl. print CSS and verbatim steps, inline-script regression on all pages).
Version 2.5.129
Batch 18, release B — the Windows client half of the batch: the registry Backup exclusions persistence wired for real (item 7), and the client half of the parser fuzz pass (item 3). Windows client only; the server stays 2.5.160, Android stays 1.29.0, Linux stays 1.25.0.
- Item 7 audit — what the code showed. The 1.37.0 pair was split-brained:
restore_critical_data() (registry → config) was alive and called at install, but backup_critical_data() (config → registry) — the writer that fills HKLM\SOFTWARE\AEAntivirus\Backup with the license key, exclusions, trusted apps and blocked apps, plus file copies of the whitelist — had zero call sites in the tree (tray.c carried only its extern declaration; the call was evidently lost in a long-ago refactor). So the Backup key only ever contained data written by 1.37-era builds. Two more gaps: the modern false-positive store (fp_exclusions.db, 2.5.73) was in neither the backup nor the restore set, and the in-memory FP table was never loaded at startup — it waited for the first cloud sync or a local mark, so exclusions did not apply in the startup window. - Decision: wire it (the writer and reader both exist in full; the invocation and the FP coverage were the missing pieces — removal would have thrown away a working restore path Austin's reinstall rule depends on).
backup_critical_data() is now called at every point the backed-up data changes: a local FP mark (fp_exclude_add), a cloud FP sync (after the store swap), tray exit (final interactive state, alongside the settings-sync flush — the lost call, restored next to its 1.37.0 extern declaration), and the install path (re-backing the just-restored state under the new build). The backup set gains fp_exclusions.db (a file copy beside whitelist.db under AEAntivirus-Backup). Startup now heals before it loads: if the live FP store is missing or zero-length and the backup copy has content, it is copied back — a live file with content is never overwritten — and fp_exclusions_reload() runs in WinMain with the other store loads, closing the startup window. The install-time fresh-restore semantics are untouched. Pinned by server/tests/client/batch18r2-backup-test.sh (17 guards: call sites, FP copy lines, restore conditions, startup ordering). - Item 3, client half — parser fuzzing. A seeded mutation harness (
server/tests/client/batch18r2-fuzz-test.c/.sh) compiled the shipped testable blocks with ASan+UBSan and ran 420,000 iterations across the sigdelta applier/parser, cmd_parse_one, the quiet-hours p_hhmm parser + qh_in_effect, and fp_normalize (extracted under a new AV25129-TESTABLE-FPNORM marker pair — comment-only change). Zero findings: no crash, no UB, every invariant held (accepted commands always carry a positive id and a bounded type; HH:MM parses only strict in-range values or −1; normalization is always bounded and terminated). The C parsers were already disciplined; the batch's one fuzz bug was the server-side webhook SSRF (release A). Eight deterministic regression pins run before the fuzz loops. - Third live drift, found by release B's pre-push verification and repaired in this publish: every manifest in the repo recorded a
sigs_sha256 that did not match the shipped signatures.db — the main/android/linux/macos manifests carried the hash of a 1,446-byte development stub (494e6f87…), beta carried a different stale hash (7b763d7d…) — while the shipped database (297,214 bytes) hashes 6813cffe…. All five manifests now record the shipped file's hash. (On the live server the runtime republishes database + manifests together, so production clients were unaffected; this was repo/deploy-source drift — precisely what release A's new drift card watches for.) The release B manifest is otherwise curated as usual: version/notes/exe hashes from this build, channel fields (sigs 38, trusted signers 1541, certs 5) from the GitHub manifest. - Build: x64 then x86 sequentially; both EXEs verified from disk (embedded version 2.5.129, item-7 marker + What's New present, no stale 2.5.128 resource string).
- Switching guide at /switch. A new public page for people moving from Norton, McAfee, AVG, or Avast: each maker's clean-uninstall steps (their own uninstaller / the Windows Settings apps list, their published removal tool named as the fallback, restart to unload drivers), a plain explanation of why only one real-time antivirus should run at a time (two engines hook the same file/process activity and end up watching each other), what to expect during the switch (short gap between products, a restart, a longer first scan, the old product's quarantine being the one thing to empty first, and the honest note that uninstalling does not cancel the old subscription), and CTAs to the free trial and /subscribe. Linked from the footer Support column and the sitemap; matching FAQ entries land on /help (switching, Comfort view, Scam Alert, loyalty, protection summary). Factual and neutral — no disparagement.
- Loyalty rewards for fixed tiers. Personal/Plus/Family/Business/Enterprise/Mobile subscriptions (Custom excluded — it has volume pricing instead) earn 5% off once they pass their 1st paid anniversary and 10% off from the 3rd, as server-managed Stripe coupons
aeav-loyalty-5 / aeav-loyalty-10 (duration forever, created via the API on first use, the 2.5.116 ensure-pattern). Tenure is counted in full months from licenses.created_at — a paid license row is created by the checkout webhook at purchase, so that is when the license first became paid (trial rows never reconcile: only active licenses are eligible). reconcileLoyaltyDiscount() runs on invoice.payment_succeeded and on customer.subscription.updated for the affected license, plus a daily loyalty-sweep job over all active paid fixed-tier licenses. One-discount discipline: the reconcile writes only when the subscription's current discount is empty or already a loyalty coupon — a referral free month (aeav-referral-free-month), a volume coupon, or any foreign/promo coupon is left untouched, and loyalty applies after a once referral coupon falls away. Idempotent: an already-correct discount means zero Stripe writes. - Comfort view (accessibility mode). A Comfort view toggle in the header of every public/account page (EN and ES chrome alike) is a plain link to
GET /comfort?on=1|0 — no JavaScript needed — which sets or clears the aecomfort cookie (about a year) and redirects back to the page you were on (same-site paths only). When the cookie is on, render() adds the comfort class to <html> plus a dedicated CSS block: larger base font size, increased line-height, and a higher-contrast palette with clearer link color. The block is served only in comfort mode, so the default chrome is unchanged. - Printable protection summary.
GET /account/protection-summary (login-required, linked from My Account) renders the signed-in account's own data only: account email, each license's plan name and status, every device (name, platform, last seen, and lifetime threats blocked from installs.lifetime_threats, the counter the heartbeat maintains), the account's active add-ons, and a generated-on date. An @media print stylesheet gives a clean paper layout — white background, black text, no nav, footer, ticker, or chat bubble. - Scam Alert newsletter. A signup form on /tools and the Threat Guide posts to
POST /api/newsletter/signup (works with JavaScript off; honeypot field plus a per-IP rate limit in the contact/chat style) and stores one row per email in the new newsletter_subs table (email UNIQUE, token, ip, created_at, unsubscribed_at) — duplicates never duplicate, and re-signing-up after unsubscribing reactivates the row with a fresh token. The monthly scam-alert job (jobRun-wrapped; checked every 15 minutes, gated to the 1st of the month and once per month via its state file, following the customer-digest pattern) composes the email from the Threat Guide's scam spotlight — the same data the guide and homepage render, three entries rotating by month — and sends via the mailer with kind scam-alert; an owner-only manual send (POST /api/admin/newsletter/send) follows the weekly-report send-now pattern. Every email carries an unsubscribe link to GET /newsletter/unsubscribe?token=…, which stamps unsubscribed_at idempotently and shows a friendly confirmation page. Admin visibility: active subscriber count and last send show at the top of the admin Email log section (from GET /api/admin/mail-log's new newsletter field). Server-only release; the Windows client stays 2.5.114.
Version 2.5.160
Batch 18, release A — the server half of the batch: a feed-health canary, an update-channel drift check, Spanish round 4 for transactional email, an on-demand backup restore drill, the trial-guard audit close-out, a bounded + admin-visible webhook delivery log, and the server half of the parser fuzz pass. Server-only release; the Windows client stays 2.5.128, Android stays 1.29.0, Linux stays 1.25.0.
- Feed-health canary (item 1). Every threat-feed pull — each MISP/TAXII/OTX source and the admin threat-intel feed — now records a row in
feed_pull_stats: when, success or failure, rows in, and the per-type extracted counts (hashes / certificates / domains / URLs / IPs) plus how many were newly added. Protection → Feed health renders every feed with its latest pull, its baseline, and warnings. The baseline is honest: the first successful pull on record (nothing is backfilled or invented, and a feed with no recorded pull reads never pulled, not OK). Two warnings exist because both are the failure shape that hurt before (the 2.5.60 ingest bug dropped hashes and cert fingerprints for months while pulls reported "ok"): type drop — a type that arrived on the baseline pull reads zero on the last two consecutive successful pulls — and stale — no successful pull within the feed's expected interval (×1.5 grace). A failed latest attempt reads ERROR with the error text verbatim. Auto-paused sources report PAUSED and are exempt — their silence is by design. Bounded to the newest 200 rows per feed. API: GET /api/admin/feed-health. - Update-channel drift check (item 2). System → Update channel (and
GET /api/admin/update-channel) compare the published manifests against the exact files the update and download routes serve — no parallel truth. Per platform (Windows / Android / Linux): every hash recorded in the manifest is recomputed from the file on disk; the installer/ copy the download page serves must be byte-identical to the updates/ copy the update channel serves; and where the embedded version is readable (the Windows installers carry it in their PE version resource) the manifest version must appear in the binary. The verdict is OK or DRIFT with each specific mismatch named — the 1.38 failure class (channel silently stale for weeks) now surfaces the hour it happens. Two live drifts found by the audit and repaired in this release's publish: all five manifests recorded a certs_sha256 that did not match the shipped certs.db (manifests corrected to the shipped file, certs_version 4→5), and server/installer/aeantivirus.apk still held Android 1.27.0 while the update channel served 1.29.0 (the installer copy is now the 1.29.0 APK, byte-identical). - Spanish round 4 — transactional email (item 4). Customers whose account language is Spanish now receive the transactional emails in Spanish: password reset, verification, expiry/renewal warning, trial warning, detection notice, gift license, win-back, monthly fleet report, and the weekly family digest (protection names included).
accounts.lang persists the customer's explicit site language choice at the moment they make it (the cookie already carried it; it now survives across devices), and every send site renders through the locale-aware template path. No separate receipt email exists — Stripe sends payment receipts — so the receipt class is covered by the expiry/trial-warning templates, which the tests pin in both locales along with a live password-reset send. Admin template overrides still win for both languages, exactly as before; the admin template editor now shows the Spanish defaults alongside the English ones. - Backup restore drill (item 5). System → Maintenance → Backups gains Run restore drill (owner-only,
POST /api/admin/backup/drill): the newest local backup is verified end-to-end on demand — SHA-256 recorded, AES-256-GCM decrypt + auth-tag check when encrypted, archive extracted to scratch, the database opened read-only, integrity_check run, and every table's row count reported. The drill shares its core with the weekly verification job (whose recorded detail strings are unchanged), records its outcome in the same history (so Server Health reflects drills too), and never touches the live database or the backup file. Off-site copies are the same files uploaded, so a passing local drill is the restore proof for both. - Trial-mint guard — audit first (item 8). The audit found the guard mostly built (2.5.132): per-IP caps (3/24h without a device fingerprint, 50/24h backstop with one), a per-fingerprint cap (3/24h), a manual blocklist, and the Trial abuse watch — on every mint path (register-mint and account/download provisioning share
ensureTrialKey). The one open gap: no per-email cap, and the fingerprint is client-supplied — a signed-in account could mint up to the loose IP backstop all day. 2.5.160 closes exactly that: each account email may mint at most 3 trials per 24 hours across all paths, and Trial abuse watch gains the by-email grouping so repeat emails are visible next to IPs and fingerprints. - Webhook delivery log — audit first (item 9). The audit found deliveries already journaled per attempt since 2.5.147 (endpoint event, HTTP status, error, attempt number) and visible to the owning customer. Missing: any bound and any admin view. 2.5.160 adds both without changing what customers see: inserts go through one journal helper that keeps each webhook's newest 500 deliveries, the table joins the data-retention pruner, and Customers → Webhook deliveries (
GET /api/admin/webhook-deliveries) shows the fleet-wide log — account, event, endpoint host, status, attempt, error — with a failures-only filter and all-time totals. - Parser fuzzing, server half (item 3). A new bounded harness (
server/tests/fuzz/fuzz-server-parsers.js: seeded PRNG, fixed corpus + mutations, reproducible seeds) ran 2,083,576 inputs across the feed-ingest extractors, the STIX parser, sigdelta build/parse/apply, and the webhook/intel URL validators — plus HTTP-level fuzz of the link checker and quiet-hours policy parsing in the release suite. One real finding, fixed: the webhook URL validator accepted IPv4-mapped IPv6 loopback literals (https://[::ffff:127.0.0.1]/…) because the URL parser canonicalizes them to hex groups (::ffff:7f00:1) that the dotted-quad mapped check never matched — ipIsPrivateOrReserved now decodes both mapped forms and judges the embedded IPv4. The intel-feed validator already refused them (pinned by probe). sigdelta's 378,816 malformed-delta rejections are its designed Error throws, all caught by its production callers; the round-trip invariant (apply∘build = identity) holds. The harness ships in the suite so the corpus keeps guarding. - Tests:
server/tests/server/v25160-batch18r1-test.js (feed canary incl. the poisoned-baseline zero-type case + a real failed-pull recording, drift endpoint cross-checked against disk ground truth, both email locales pinned on templates + live sends, drill on a seeded backup incl. live-DB-untouched, the per-email cap with a fresh-email control, the 500-row journal bound through a real test-send, fuzz-in-suite, inline-script regression on all pages).
Version 2.5.128
Batch 17, release E — update rollback grows to cover full-installer upgrades. Windows client only; the server stays 2.5.159, Android stays 1.29.0, Linux stays 1.25.0.
- Installer upgrades get the Batch 15 rollback watch (item 7). Batch 15's rollback covered only the auto-update single-exe path, because the updater stub was the piece that retained the outgoing build. A full Setup run had no stub — it just replaced the binary. Now, inside
install_core_ex (the one core behind the GUI, console, and silent installers), the installer retains the outgoing installed build as AEAntivirus.prev.exe before the staged swap, and only after the swap actually lands writes the same update_rollback.ini pending record — with from_version read from the retained build's own version resource and a via: installer marker — clears stale markers, and schedules the same AEAntivirusRollbackCheck one-time task, pointed at the installed exe (--rollback-check). From there everything is literally the auto-update machinery: the started marker at first heartbeat, the health marker at first server contact, the unchanged rb_decide rules (never-started past the 15-minute window → the previous build is restored and restarted; started-but-unproven → kept, never rolled back), and the restored build journals the same update_rollback event an admin sees for auto-updates. - The watch is armed only for a build that is actually installed. If the file swap is deferred to the next reboot (files in use) or fails outright, the installer disarms again — retained copy removed, no record written — and the log states plainly that the upgrade has no rollback coverage, the same honesty rule the auto-update path uses when retention fails. A fresh install retains nothing and arms nothing. Rollback still keeps exactly one previous build, not a history.
- One supporting change: the running build now settles a concluded watch once per process at heartbeat time (the decision runs after the started marker is written, so a live build can never roll itself back). Installer upgrades have no stub monitor; this plus the scheduled task gives them the same two deciders, and it also closes the auto-update corner where a dead stub monitor and a missed task left a watch pending forever.
- Tests:
server/tests/client/installer-rollback-test.sh (29 checks — arm API + shared path derivation, retention-before-swap ordering in the installer, arm/disarm placement on the landed/deferred/failed branches, backstop target, heartbeat settle, and the Batch 15 suite re-run green). Both architectures rebuilt and verified from disk against src/cloud.h.
- Mobile responsiveness pass across the whole site. Every served surface was audited at phone (360px / 390–430px) and tablet (768px) widths and the fixes are additive CSS in the two shared style blocks — desktop rendering above 1024px is unchanged. In the public chrome (
render()): the top nav is now one bounded horizontally-scrolling row under the brand at ≤860px (previously the growing nav — 6–7 links plus the EN/ES toggle — wrapped into a tall stack under the sticky header); page side padding steps down 32px → 20px → 16px; the footer grid goes 6 columns → 2 at ≤860px → 1 at ≤560px with taller tap targets on its links; form controls (select, textarea, number/url/tel/search/date inputs, and the chat widget's textarea via a targeted !important) get 16px text at ≤640px so iOS Safari no longer auto-zooms on focus; table cells wrap long words, account device rows (.instrow) wrap with breakable names, hero CTAs stretch full-row, and the chat bubble insets to 12px. An overflow-x: clip on body is a last-resort net only — wide content is contained by the per-element rules (tables have scrolled inside their cards since the earlier 640px rule; the comparison and ladder tables on /subscribe and /business ride on it, and /status already wraps its table in a scroll card). In the admin console: the 1.37.3 bottom tab bar squeezed all 10 tabs with flex:1 (~36px each at 360px, labels wrapping) — it now scrolls sideways as one row of ≥62px tabs, hidden scrollbar; data tables scroll inside their cards at every phone width (the display:block; overflow-x:auto treatment previously applied only under 480px), stat cards go two-up, filter bars cap their inputs at container width, and the content column's side padding tightens. The reseller portal shares the public chrome and is covered by the same rules; the standalone vault page already carried its own viewport + max-width layout. Verification is static-plus-served (no layout engine exists in this repo): v25128-responsive-test.js boots a real server and asserts the viewport meta, the 2.5.128 rule signatures in the served chrome and admin CSS, table-scroll coverage, newest-3 homepage structure, and that every audited page still returns 200. Server-only release; the Windows client stays 2.5.114.
Version 2.5.159
Batch 17, release A — the server half of the batch: live-data ML evaluation in the admin panel, a real Verify for reseller domains, the paid-mismatch list, an admin performance pass, and the server side of Android false-positive sync (the Android half ships in 1.29.0). Server-only release; the Windows client stays 2.5.127, Android stays 1.28.0.
- ML evaluation on live data (item 2). The admin training panel (Detections → Trained ML model) gains Run evaluation on live data: the Batch 15 corpus evaluation — labeled verdict corpus, dedupe, time-ordered holdout, baseline vs full-corpus candidate, ship rule — runs in-process against this server's live database and renders the numbers verbatim, including the ship/no-ship decision. It is read-only (it never changes the shipped model), owner-only, audit-logged, and bounded: a corpus beyond the in-panel limits returns the counts with a pointer at the standalone harness (
server/tests/ml/corpus-eval.js) instead of wedging the server. An empty or thin corpus reads as exactly that — the same NOT EVALUABLE wording the harness prints. The in-panel numbers were verified identical to the harness's on the same database (corpus totals, holdout metrics, decision). - Reseller domain Verify (item 5). Next to the existing Check (still the only action that verifies a domain for branding), resellers and the admin list gain Verify: the server walks the domain's DNS (CNAME chain or final addresses — does it lead here?) and then makes a real TLS handshake for the name, and reports in plain words — *Points here — HTTPS working*, *Points somewhere else* (naming where), *Points here, but HTTPS is not working* (with the usual cause: the domain also has to be added in the hosting dashboard for a certificate to exist), or *Does not resolve*. All lookups are time-bounded; Verify changes nothing.
- Paid plan/platform mismatches, listed (item 6). The Licenses page gains a Paid plan / platform mismatches card: the
license.platform_mismatch audit entries the 2.5.158 boot check writes are read back, deduped per license (masked key, platform vs plan, first/last seen, report count), so each one can be decided from a list instead of the audit log. Paid licenses are still never changed automatically; the card says where its data comes from (the retained audit log) and shows an honest empty state. - Admin performance pass (item 9). The DATA_DIR JSON stores (queued commands, policy overrides, account policies, threat list) were re-read from disk and re-parsed on every call — the command store alone is loaded on every device check-in and every fleet-list request (a ~0.9 MB store cost ~6 ms of read+parse per call on the bench VM, growing with fleet size). Loads are now mtime-cached and saves refresh the cache; the caller contract (mutate, then save) is unchanged. The audit viewer endpoint now scans the file log in growing windows and stops once a page is filled, reporting
total_exact so the count stays honest ("N+ matching" when the window stopped early). Before/after on the seeded 5,000-install / 50,000-detection bench (this VM, medians of 7): endpoint medians were within noise at this scale (fleet list 61.0 → 64.2 ms; the 1.4 MB / 2,000-row payload dominates it), while the per-call store re-parse (~6.8 ms across the three stores on the fleet/heartbeat paths) and the audit viewer's full-log parse are eliminated as growth terms — they no longer scale with fleet size or log length. No verdict is claimed about the browser-side memory Austin reported; that check remains his on-screen one. - Android false positives (item 1, server half). Verified end-to-end with real requests: an Android install's FP verdict lands in the same per-device exclusion store as every other platform, the device pull returns it, and the admin exclusions viewer shows the Android rows (nothing was missing server-side; the app-side reporting ships in Android 1.29.0).
Version 2.5.127
Batch 16, release A — the Windows client reorganizes around how people actually look for things. Nothing was removed: a feature inventory taken from the app itself (41 sidebar pages, 130+ settings, every tray-menu entry) is pinned by a guard test (server/tests/client/batch16r1-test.sh), and every inventoried feature is still reachable in at most two clicks. Client-only release; the server stays 2.5.158, Android stays 1.27.0.
- Grouped sidebar. The flat list of 40 pages is now grouped by job — Overview, Scan, Protection, Network, Privacy, Family & Safety, Tools, Performance, Help — with Status first and Settings pinned at the bottom, as before. The inventory also caught two pre-existing defects, both fixed: a duplicated block created five sidebar buttons twice (the copies sat outside the scroller), and the tray menu's Password vault entry could never open the password vault (it shared an ID with the encrypted vault and always opened that instead).
- "What moved" panel + guided tour. After updating, a one-time panel lists every page's old and new home; it never appears on a fresh install and never twice. A 9-step guided tour (from the panel, or Status → "Show me around") walks the new groups page by page.
- Settings search. The Settings window has a search box: type part of a setting's name, click a result, and the app switches tab, scrolls the setting into view, and flashes it. A generated registry (127 entries) covers every settings control and is guard-tested against the settings builders both ways.
- Comfort view. Settings → General → Appearance: larger text everywhere and a roomier sidebar, applied at restart, carried by settings sync like any other preference.
- Recovery & Support on System Tools. The Shredder, the Ransomware Rollback vault, the AI Skills Checker, and offline update import gain one-click homes in a new right-hand column, next to two new tools: Create support bundle (one text file with versions, license *state* — never the key — settings, and recent events; save-only, because the bug-report channel takes text only) and settings export/import to a file, reusing the exact sync blob so fleet-managed keys are excluded on export and skipped — and counted — on import.
- Regrouped tray menu. Quick Scan, a new Full Scan entry, Pause, Focus mode, Check for updates, and Settings on top; everything else under "More tools". Every prior entry survives (the inventory maps old → new one by one).
- Per-page help. Every page's ? button (top right) opens the Feature Guide at that page's own section; seven guide sections were added so pages that lacked one (Kernel Guard, Boot Guard, Ransomware Rollback, and friends) anchor properly. The guide was deduplicated to one section per feature.
- Safer destructive actions. Reset statistics now requires typing RESET (the button stays disabled until the word matches exactly).
- Keyboard & screen readers. Buttons are in the tab order with a visible focus outline; the protection toggle switches carry accessible names. (Screen-reader verification itself needs a pass with NVDA on a real desktop.)
- Footer completed: every public page is now linked from the footer. The shared footer (rendered by
render() on every public page, English and Spanish chrome alike) was audited against the actual route table and rebuilt into labeled columns — Products (Download, Pricing, Free trial, Features, Paid add-ons, AE for Business, Reseller program), Free tools (All free tools, Link checker, Scam message checker, Security checkup, Threat Guide), Support (Help & FAQ, User guide, Contact us, About us, System status, Changelog), Legal (Terms of Service, Privacy Policy, License Agreement, Cookie Policy), and Account (Sign in, Create account, My Account). Newly linked pages that were previously reachable only by nav, in-page links, or direct URL: /business (2.5.123) and /security-checkup (2.5.125); the old single "For business" link to /reseller is now labeled Reseller program so the two business pages are distinguishable. The foot grid widens from five tracks to six to hold the added Free tools column; the brand blurb and the "Proudly designed and coded in the USA / Copyright 2026 Mott Enterprises" bottom bar are unchanged. Deliberately not linked: /admin and /admin-login (unlinked by design), all /api/* endpoints, the transactional landings (/success, addon success, /verify-email), the file-download URLs (/download/file/*, the setup EXE), and the login/add-on-gated app pages (/vault, /mobile). /licenses needs no entry of its own — it is a redirect alias for /account, which is linked. A verification pass requested every footer href against a booted server: the footer link set is identical on the homepage, /subscribe, and the tool pages, and every link returns HTTP 200 for a logged-out visitor (account pages render or redirect to /login, never 404). Server-only release; the Windows client stays 2.5.114.
Version 2.5.158
Austin's rule, enforced end to end: Android installs only get the Mobile plan. Server-only release; the Windows client stays 2.5.126, Android stays 1.27.0.
- A trial takes the plan of its platform. A trial issued for Android is a Mobile plan trial from the start. (Before this release the trial kept the generic tier, which My Account displayed as "Personal" — an Android phone on a "Personal plan" is what prompted the fix.) Trials with no platform stamp yet (legacy rows) take the platform and plan of the first device that activates them, at registration or heartbeat; page-provisioned trials are Windows trials. Trials already on file with an Android stamp — or an Android device registered under them — were flipped to Mobile at the first boot of this release.
- Cross-platform keys stay refused (registration enforcement since 1.38.7, unchanged): a key activates only installs on its own platform, with an error naming both platforms; legacy keys with no platform stamp fall back to their plan's platform.
- Paid mismatches are reported, never flipped. A paid license whose plan and platform disagree is left exactly as it is and listed (key masked) in the boot log and the audit log (
license.platform_mismatch) for a human decision. - My Account speaks device. The license card's wording follows the license's platform — an Android license now reads "Devices using this key" and talks about your Android device, not a "PC"; unnamed Android installs read "Unnamed phone". Spanish strings updated to match.
Version 2.5.126
Batch 15, release C — the client consumes signature deltas, gains the update rollback safety net, and enforces family quiet hours. The server stays 2.5.157.
- Signature deltas (item 5, client half). When the manifest offers a newer signature database, the client first walks the server's AESIGDELTA chain from the version it holds: each delta is downloaded, its own SHA-256 checked, applied to a working copy by a C engine that mirrors the server's reference implementation byte-for-byte (proven against the server's own outputs and the shipped chain artifacts in the test suite), and the rebuilt database is hash-checked per hop and finally against the manifest's full-database hash. Only a fully verified rebuild is swapped in and reloaded. Any failure — a chain gap, a bad hash, a line that doesn't apply — is logged with its reason and falls back to the existing full download, unchanged. Devices whose signature version predates the chain simply take the full download as before.
- Update rollback (item 6). Every engine update now retains the outgoing build (one generation) beside the new one and opens a pending record before the swap. The new build proves itself with two markers: a started marker the moment its engine runs, and a health marker at its first successful server contact. The updater stub — a copy of the *old* binary, so it survives a new binary that cannot start at all — stays behind as monitor, with a one-time scheduled task as backstop across reboots. Verdicts: healthy → watch closes, previous build kept until the next update; started but unproven past the 15-minute window (the normal case: the PC is offline) → kept, never rolled back; never started → the retained build is restored, restarted, and the restored build journals an
update_rollback security event. Scope, stated plainly: this covers the auto-update single-exe path; full-installer upgrades rely on the Repair path (Batch 13). If the previous build cannot be retained (disk error), the update proceeds and the log says it has no rollback coverage. - Family quiet hours (item 10, client half). The heartbeat now carries the per-device quiet-hours policy a parent set on /family (enabled, start, end, days); the client stores it and parental controls enforce it in the web filter: during the window, the entertainment + social categories (social, gaming, streaming) are blocked with the reason
quiet-hours — everything else (school, news, banking) keeps working. Overnight spans belong to the day they start (a 21:00–07:00 Mon–Fri window covers Saturday 02:00 but not Saturday 21:00), the window is start-inclusive/end-exclusive, and malformed policy fails open. Blocks journal quiet_hours_block security events (first block per domain per 10 minutes; the parental log keeps every block) which the family digest counts — they never touch threat counters or detections. The journal's event-type field widened 16 → 24 characters so the 17-character type is stored whole end-to-end. - No model change (item 4). The retrained-extraction candidate is not shipped: per the 2.5.157 harness decision, the build environment holds zero usable verdict samples and a synthetic corpus is not a ship basis. The client model and the 116-signal scoring are exactly as in 2.5.125.
- Sweep fixes (item 1, client half). The six main-window button-ID collisions the full-suite sweep found (IDs 786–791 renumbered off the navigation band) and the stale button labels it flagged are fixed.
- Admin: Support inbox (Customers → Support inbox). One queue over the
support_messages store (2.5.125): the chat bubble (source='chat') and — new in this release — the /contact form, which now also mirrors every message here (source='contact') while its own contact_messages store and email behavior stay exactly as before. GET /api/admin/support-messages?source=&status=&archived=1 (archived rows are hidden unless requested), with per-status counts. Actions (owner-only): Mark read and Archive (POST /api/admin/support-messages/read|archive), and Reply (POST .../reply): the reply is sent through the mailer (kind='support-reply'), stored on the row (reply_body, replied_at), and flips the status to replied. With SMTP unconfigured the reply returns a visible 400 and nothing is marked sent — no silent pretend-replies. All actions audit-logged; the section follows the independent-loader pattern. - Admin: Email delivery log (System → Email log). The mailer is instrumented centrally:
sendMail is now a wrapper over the original implementation that journals every attempt into a new mail_log table (kind from opts.kind — contact/chat/support-reply sends are tagged, everything else defaults to email — recipient, subject, ok, short error, timestamp), including 'SMTP not configured' rejections, so configuration problems are visible in the log too. Journaling sits inside try/catch and can never change a send's outcome (the wrapper resolves/rejects exactly as before). GET /api/admin/mail-log?kind=&status=. Retention: rows older than 90 days are pruned opportunistically by the wrapper (at most once per 24h) and via Data cleanup. - Admin: Scheduled jobs viewer (System → Scheduled jobs). An honest inventory of the server's 18 real periodic jobs (weekly report, customer digest, dunning, trial nurture, insights digests, payouts, alert rules, health alerts, site monitor, cert monitor, threat-intel sync, MISP sync, ML auto-train, breach checks, retention, file cleanup, auto backup, quarantine cleanup), each scheduler entry wrapped in a minimal
jobRun() journal (job_runs: job, trigger schedule/manual/startup, started, finished, ok, detail). GET /api/admin/jobs shows last run + outcome and a next-expected time derived from the last run plus the job's check interval (gated jobs still act only on their internal schedule, which the row states). Run Now (POST /api/admin/jobs/run, owner-only) exists only for jobs that are safe and idempotent by hand — threat-intel sync, MISP sync, ML weights, breach checks, retention, file cleanup, auto backup, quarantine cleanup. Email-sending, money-moving, and alert-firing jobs are listed without a button, with the one-line reason on the row. - Admin: Announcement banners (System → Banners). CRUD over a new
site_banners table (text, optional link URL + label, style info/warning/success, starts_at, ends_at, enabled) via GET/POST /api/admin/banners + /update + /delete (owner-only mutations, audited). At most one banner renders at a time — the newest enabled banner inside its window — prepended at the top of /, /account, and /subscribe only. The admin section previews whichever banner is live right now; not-yet-started, expired, and disabled banners never render, and two in-window banners never stack. - Admin: Data cleanup (System → Data cleanup).
GET /api/admin/cleanup shows row counts and oldest-entry age for exactly the prunable stores — detections, mail_log, login_attempts, verdict_reports, job_runs, and archived support_messages — plus the audit log's line count as information only (it is a file, audit.log, which auditLog() already rotates to the newest 5000 lines; there is deliberately no prune action for it). POST /api/admin/cleanup/prune accepts only a registry table and a keep window of 30/90/180 days (anything else is rejected), and POST /api/admin/cleanup/vacuum runs VACUUM with a "brief write lock" note in the UI. Every action is audit-logged. Hard rule, enforced by the registry: licenses, accounts, installs, and payment/subscription data are never listed and never prunable here. Server-only release; the Windows client stays 2.5.114.
Version 2.5.157
Batch 15, release B — signature deltas, family quiet hours (server half), the ML corpus harness, and the extension's domain feed. Server-only; the Windows client stays 2.5.125.
- Signature delta publishing (item 5, server half). Every signature publish now also emits a small AESIGDELTA artifact capturing the exact line difference from the previous version, chained in
updates/sigs-chain.json (last 8 kept) with the delta's own hash and the resulting full-database SHA-256. Each delta is self-checked before it is published — applied back onto the previous file and hash-compared — so a bad delta can never ship; clients that don't use deltas keep the full download, untouched. Client 2.5.126 consumes the chain. - Family quiet hours (item 10, server half). Parents set a per-device quiet-hours window (enabled, start, end, days) on the /family page while signed in — a new account-authenticated endpoint writes it as a per-install policy override, and it reaches the device through the normal heartbeat policy fields. The weekly family digest gains a per-device "Quiet hours: N blocks" line fed by the device's
quiet_hours_block events (the line is omitted at zero, and those events never touch any threat counter). The security-event type cap was widened to 24 characters so the event type is stored whole. The client enforcement ships in Client 2.5.126. - ML verdict corpus + evaluation harness (item 4). New standalone tooling (
server/tests/ml/corpus-eval.js) assembles the labeled corpus under exact rules — verdict-log groups plus detection verdicts with canonical signal tags; public link/scam checker reports excluded (different feature space, counted and reported); triage states are workflow, never labels; identical samples dedupe, conflicting labels drop as ambiguous — then trains the existing gradient-boosting pipeline on verdict-log data alone vs the union and scores both on a time-ordered holdout, parsing the real MLMODEL text artifact back for scoring. Decision: extraction is NOT changed in this release. The only database reachable from the build environment holds zero verdict history (the real corpus lives on the production server), and a synthetic corpus — which the harness also runs, proving the mechanics end to end — is not a basis for changing what production training learns from. Run the harness against the production database to get the real holdout numbers; the ship rule (candidate FP rate not regressed, detection improved-or-held) is encoded in the tool. - Public threat-domain feed (items 7/8, server bits).
GET /api/threat-domains serves the compact union of threat-feed and web-filter domains (bare hostnames, sorted, cacheable for an hour). It exists because the browser extension holds no install token — the existing feed endpoints all require one — and it serves only the blocklist itself; Android keeps using the authenticated threat feed unchanged. - Privacy policy now discloses that threat/false-positive verdicts (the technical signal tags + timestamps, never file contents) are used to train the on-server threat model.
Version 2.5.156
Batch 15, release A — test-suite honesty, scale, badge accuracy, and SIEM forwarding. Server-only; the Windows client stays 2.5.125.
- Threat feeds deliver everything again (real bug, fixed). The feed-ingest path silently dropped file hashes and SSLBL certificate fingerprints from every pull since 2.5.60 — ThreatFox, MalwareBazaar, TAXII, and OTX hashes never reached the signature store, and URL attribution metadata was skipped (URLs and IPs survived, which masked it). The cause was a type guard substituting an empty list for the Set the extractors return. Fixed at the ingest boundary; feed intelligence is complete again.
- Module badges tell the truth (item 3). The fleet console painted every device all-green by default: a device that never reported its module state showed protection running, an omitted security score snapped back to a perfect 100, and nothing ever looked stale. Heartbeats now keep the last reported state when fields are absent, the server stamps when state was last reported, devices that never reported show grey n/a badges, and state older than 24 hours renders greyed as last-known with its age. The audit traced every module (real-time, firewall, web protection, advanced, USB, pause) from engine getter to admin render; the Windows client already reports real values, and Android/Linux — which report none — now honestly show n/a instead of green.
- Syslog / SIEM forwarding, completed (item 9). The admin-configured forwarder gains per-class toggles (detections and tamper events on by default; the new admin-audit stream off by default), a bounded send queue with delivery stats (sent / dropped / queued / last error / last sent), and a Send test event button that exercises the real path. Forwarding can never block or fail a request, and the legacy environment-variable detection path now stands down when the admin target is active, so detections are forwarded exactly once. The privacy policy discloses the forwarding (off by default; device identifiers, threat names, timestamps — never file contents or browsing history).
- Scale, measured (item 2). A seed + benchmark utility ships in the test tree (5,000 devices, 50,000 detections, full triage/tamper/exclusion spread). Two indexes were missing and are now created at boot (open-detections, heartbeat order); the fleet list now reports its total device count and says plainly when only the newest 2,000 are shown instead of silently truncating. On this VM the board and list endpoints hold at ~26–74 ms medians at that scale; the wins are structural (index coverage + honesty), verified before/after in the release notes.
- Test-suite honesty pass (item 1). Every suite in the tree passes or is explicitly accounted for — none retired, none silently skipped. Stale suites were updated to assert current behavior with computed expectations (never new hard-coded versions or counts); the sweep also caught the feed bug above, six colliding button IDs in the Windows UI, and stale button labels — the client-side fixes ship with Client 2.5.126. The ML training counter also refreshes immediately when verdicts land instead of lagging up to five minutes.
Version 2.5.125
Batch 14, release B — the client half. Everything is on-device except where noted.
- Quarantine encryption at rest (item 1 — verified first). The audit confirmed quarantined files were stored raw (byte-identical copies under a
.quar name). New quarantines are now encrypted in place with AES-256-GCM (the codebase's existing AEAVENC1 frame from the folder-encryption feature) under a dedicated per-machine key protected by DPAPI machine scope — machine scope because the SYSTEM service quarantines files the interactive user's tray and CLI must be able to restore, fetch, or detonate; a user-scope key would be unreadable across that boundary. The key loader fails closed (an existing key that cannot be unprotected is never overwritten — overwriting would orphan every stored file). Stored files are recognized by their AEAVENC1 frame header, so restore, FP-restore, delete, admin fetch_quarantine, and sandbox detonation all handle legacy raw and new encrypted records transparently; the admin fetch decrypts before upload so the console still receives the real sample. A one-time sweep at service start encrypts pre-existing raw quarantines in both user and service vault locations and journals the count. Zero-byte quarantines stay raw (nothing to protect). The same pass fixed a latent folder-encryption bug: the folder key was protected *with* DPAPI entropy but unprotected *without* it, so a stored key could never reload after restart — and the fall-through generated a fresh key over the old one, orphaning prior encrypted files. Unprotect now passes the same entropy. - USB auto-scan modes (item 2 — arrival scanning already existed). The engine already root-scanned new removable drives on arrival; this release adds the mode on top: off / prompt (default) / auto, set locally in Settings or by the new fleet policy
policy_usb_autoscan (which ships in this release's server half, following the policy_usb_block precedent — no separate server bump). Prompt uses a tray handshake: the dialog waits while Focus mode is active, an unanswered prompt expires after 10 minutes as a skip (prompt mode never scans unasked), and every arrival, mode, and outcome is journaled. Threats found follow the normal detection flow. - Wi-Fi security grading (item 3). The existing Wi-Fi security poll (WlanAPI — already linked for the 1.5.0 module) now grades the connected network from its negotiated auth/cipher: Unprotected (no cipher), Weak (WEP family, shared-key), Fair (WPA-era, TKIP, or OWE), Strong (AES with WPA2/WPA3), Unknown (never dressed up). The grade shows on the tray Network page next to the SSID; when a known network comes back with a weaker grade than before, the client journals a drop warning ("this network changed") — the evil-twin signal.
- Home network inventory (item 4, client half). The LAN scanner (ARP table + throttled SendARP probing of the client's own /24 only — never beyond the local subnet) now resolves a vendor from a baked-in OUI table (unlisted = "Unknown"; a locally-administered MAC is reported as a randomized/private address, which is a fact about the address, not a guess), keeps first/last seen per device, lets you name devices in the list, fires the
lan_new_device security event on first sighting, and reports the sanitized list in the heartbeat for the admin console view shipped in Server 2.5.155. - Focus mode (item 5 — the DND queue and silent mode already existed). New on top: true-fullscreen detection drives a real queue — while a game or video owns the screen, detection dialogs and balloons queue and appear in order when focus ends; critical (ransomware/vault) events break through immediately; scheduled scans postpone until focus ends (journaled, with catch-up scans also waiting); USB prompts hold. The deferral decision is a pure function (
class × focus state → show/queue/break-through), unit-tested, with a bounded FIFO queue. Journal records focus start/end; the tray menu shows the toggle (default ON) and an "active" indicator. - Secure shredder (item 6 — the engine already existed since 1.4.7). Added: 1-pass / 3-pass choice in the UI (1 is the offered default; invalid selections fall back to 3), filename obfuscation before deletion (the file is renamed to 16 random hex characters in the same folder first, so the original name doesn't linger in directory entries), per-file result counts in the completion report, and the confirmation now states the SSD caveat verbatim: *"On SSDs, wear leveling means no overwrite can be guaranteed — the drive may keep copies of some blocks in spare cells the software cannot reach."* The Explorer right-click registration is unchanged from the existing pattern.
- Settings sync (item 7, client half — the sync channel existed since 1.46.0 for exclusions). The client now syncs its settings and exclusions through the existing
/api/settings-sync channel extended in Server 2.5.155: a whitelist of ~44 user-preference keys (identity and secret keys — license key, install id, token, server URL — are never in the blob, by construction), pushed debounced after changes and at shutdown. On a fresh install with a restorable blob, settings apply once; the /account "Restore on next check-in" flag triggers a one-time apply even over existing settings, with a tray note. Policy wins: any key a fleet policy manages on that device is excluded from both the upload and the apply, mechanically, via the same *_managed markers the policy engine sets. - Silent deployment (item 8, installer half). The installer accepts
/silent (and --silent) and reads aeantivirus-setup.ini beside the exe — exactly the four keys the admin guide's downloadable example documents (server_url, license_key, accept_tos, accept_license), pinned by tests on both sides. Fully unattended: service installed and started, registration attempted, no window or prompt ever. Deterministic exit codes: 0 ok (including reboot-pending), 2 not admin, 3 ini missing/unreadable, 4 terms not accepted, 5 license key missing, 6 install failed — every failure logged with its reason; a missing/invalid key fails fast instead of hanging on a prompt.
- Help Center search.
/help opens with a search box: typing filters the FAQ entries live (every entry carries a stable hq-N anchor), matching groups auto-open, non-matching ones hide, and a result count ("3 topics match") or a "no matches — try the contact form" message shows. 67 questions across 5 sections are searchable; with JavaScript off the full page renders exactly as before — every group and answer is server-rendered. - Homepage, Subscribe and Help in Spanish.
?lang=es (or the aelang cookie, set for 30 days by following any language link) switches the whole public chrome plus /, /subscribe, and /help to natural Spanish copy (views/home-es.js, views/help-es.js, TIER_FEATURES_ES + the subscribe view's ES dictionary). An EN/ES toggle sits in the nav of every public page and is sticky across the three translated pages. All dynamic data is shared with English — prices come from the same tier/pricing code at request time, versions from the same manifest source, live stats from the same endpoint — nothing is duplicated or hardcoded. Scope is exactly those three pages: every other page keeps English chrome with the toggle present; the legal documents (/terms, /privacy, /license) deliberately stay English-only and, when the visitor's language is Spanish, open with a Spanish note saying the binding text is the English version and offering the chat/contact route for a Spanish explanation. - Real service status page.
/status is rebuilt on live self-measurement (cached 60s): Website & API (this server answering, with its version and uptime), Update channel (the on-disk update manifest parses and shows its per-platform versions; a missing/corrupt manifest reads Down), Database (a trivial query), Email delivery (SMTP configured → "Configured", otherwise honestly "Not configured"), and Threat feeds (last update time taken from the threats file's own updated field with file mtime fallback; when no signal exists the indicator reads Unknown — never a fake green). A new status_incidents store powers incident notes: the admin posts/resolves incidents from System → Service incidents (GET/POST /api/admin/status-incidents + /resolve — deliberately not /api/admin/incidents, which is the older XDR incidents API over a different table, owner-only, audited); active incidents are pinned prominently on /status and resolved ones move to history. - "Chat with us" bubble. Every public page (admin panel excluded) gets a chat-style bubble: name, email, message. Submissions land in a new
support_messages table (source='chat', status lifecycle new/read/replied/archived, reply fields reserved for the admin inbox shipping in 2.5.126) and are also emailed to the support inbox when SMTP is configured — same pattern as the contact form, including a honeypot field (filled bots get a silent success and write nothing) and a 5-per-hour per-IP cap (POST /api/support/chat). The visitor sees a clear "we reply by email" confirmation. - Personal Security Checkup.
/security-checkup (linked from /tools and the Threat Guide) asks 10 plain questions about a visitor's setup and habits and scores them entirely in the browser — nothing is sent or stored, and the page says so. Four result bands (Well protected / Good — a few gaps / Exposed in places / High risk), each answer carries a tip that links to the specific Threat Guide anchor or site page that fixes it. Server-only release; the Windows client stays 2.5.114.
Version 2.5.155
Batch 14 release A (server-only; the Windows client stays 2.5.124, its halves of items 4/7/8 ship in Client 2.5.125):
- macOS platform recognition (item 9).
INSTALL_PLATFORMS = windows/android/linux/macos; validPlatform() validates against it (register, heartbeat, netstats, bug-report), PLATFORM_ALIASES gains mac/osx/darwin → macos, platformLabel knows "macOS", and checkout (Custom tier), the Stripe webhook platform scoping, and admin bulk-keys all accept macos. A macOS install now stays macOS across heartbeats and a macOS-scoped key refuses a Windows client with the correct-direction message. There is still no signed macOS build: PLATFORMS (the download set) has no macos entry, the download page shows an honest placeholder card with no file link, and /status names the state ("macOS: no signed build yet"). The only macOS manifest remains the source-only one (its notes say "no signed binary"). Suite: v25155-batch14r1-test.js; v25138 alias-map pin updated. - Fleet N+1 killed (item 10a).
adminInstallsList (and the account-detail installs map) resolved each row's license with getLicense(key) — one query per row. Both now use licensesByKey(), a single IN (...) lookup whose map semantics are identical (exact key match, null when absent). The suite pins the fleet endpoint's JSON field-by-field for a seeded fleet (active/past_due/unknown-key/null-key rows, plus the computed fields) so the refactor cannot drift behavior. - Audit-log viewer (item 10b). The DB
audit_log backs the admin Audit Log section (the file audit.log backs the separate /api/admin/audit feed). GET /api/admin/audit-log gains structured filters — actor, action (prefix), target, from/to date range — alongside the existing free-text q; total now counts the FILTERED set, and pagination is bounded (limit 1–200, default 50). The admin section gains the filter inputs + a pager. Access unchanged and verified: owner + Manager; the path stays off the Support allowlist (a Support helper gets 403). - Settings sync, extended — not duplicated (item 7, server half). Verification found the system already existed:
settings_sync + POST/GET /api/settings-sync since 1.46.0, syncing exclusions + toggles from poweruser.c. So no new store was built. The existing endpoints are extended: the silent 64 KB truncation (which could store corrupt half-JSON) is replaced by a 256 KB cap with a clean 413; rows gain schema_version and restore_requested_at; POST honors ack_restore; GET reports schema_version + restore_requested. The shipping client's contract ({settings} up, {ok, settings, updated_at} down) is unchanged — v1460 still passes its sync checks. /account gains a Settings backup card (More tab: last-synced time + "Restore on next check-in") backed by GET /api/account/settings-sync + POST /api/account/settings-sync/restore. The restore flag is honored exactly once (client applies, then acks). Privacy Policy §1 and ToS §6 now disclose the settings backup and the home-network inventory. - LAN inventory storage (item 4, server half). The heartbeat accepts
lan_devices, sanitized by the new pure sanitizeLanDevices() in lib/upgrades-util.js (strict MAC shape, dedupe, cap 128 entries, name/vendor capped at 64 chars, control characters stripped) and stored as the latest snapshot per install (install_lan_devices, replaced — not appended — each heartbeat). The admin install detail returns it and renders a read-only "Home network devices" table. New-device sightings travel as lan_new_device security events; the suite proves they create no detection and move neither the global counter nor lifetime threats. - Silent deployment docs (item 8, server half). The admin guide gains a Deployment section: download the example config from
GET /api/admin/deployment-ini (admin-only), set the license key, place it beside the installer, run with /silent. The ini keys are the contract — server_url, license_key, accept_tos, accept_license — named in one place (SETUP_INI_KEYS); Client 2.5.125's installer implements exactly these keys and its suite cross-checks them against this served example. - Tests:
v25155-batch14r1-test.js — 82/82 over real HTTP (macOS register/heartbeat/bug-report/bulk-keys + honest surfaces, fleet JSON pin + source guards, audit filters/pagination/roles, settings-sync cross-account isolation + 256 KB cap + restore-flag round trip, LAN caps/sanitization + counter pins, ini contract, inline-script regression).
Version 2.5.124
Batch 13 release C (Windows client; the server stays 2.5.154 apart from the small fleet-policy field noted in item 5):
- Ransomware rollback, integrated (item 3). Mapping found the client ALREADY had a per-file versioning system — the rollback versioning module (2.3.0): its own watcher over Documents/Pictures/Desktop, a rollback store under the program data dir, oldest-first pruning at a configurable cap (the store-cap setting, default 2 GB), and a restore dialog on the Ransomware Rollback page. Building the spec's separate vault would have duplicated it, so item 3 integrates instead: (a) the protected-folder watcher now feeds the same store —
pfold_on_write() calls the module's new capture function, so files in protected folders are versioned alongside the user folders, one store / one index / one restore UI, and no new watcher was added; (b) every captured version records its SHA-256 in a sidecar file (the version enumerator skips sidecars); (c) restores are verified — the restore function was rewritten from copy-straight-over (a failed copy could truncate the user's file) to the verified pattern: copy to a temp file in the destination directory, size + SHA-256 verified, current file renamed aside, temp renamed in, aside dropped only after success; (d) folder restore — a new folder-restore function + a "Restore folder (newest)" button in the restore dialog restores the newest version of every file under the selected file's folder; (e) the per-file capture cap rises 64 MB → 100 MB (spec), and Settings gains the store cap field. Snapshot-timing truth (unchanged from 2.3.0): watchers fire AFTER a change lands, so the store holds the version the app saw at each change event — the rollback target when a LATER change is destructive; a file already damaged before it was ever seen has no clean copy. Store copies are the files as they were on disk — not encrypted — and the help copy says so. - Repair install (item 4). When Setup (GUI or console) detects an existing install — installed program file, registered service, or an identity-carrying config (
install_detect_existing()) — it offers Repair: install_core_ex(..., repair=1) reinstalls all program binaries (the usual stage-then-replace), re-registers the service, and explicitly re-registers the 2.5.123 Security Center presence. Preserved, never touched by the repair path: the license (config + registry backup), all settings (config.ini), local exclusions, quarantine contents and records, statistics, and the rollback vault. Repair asks for no license key, runs no account/activation step, resets no counters, and skips the competing-AV disable step and the install-mode switch. No existing install → no Repair offer. Real-machine install/repair verification is Austin's, as with all installer work. - Restore point before remediation (item 5).
handle_detection_file() — the single choke point for scan-driven quarantine/delete and real-time quarantine — first calls rp_before_remediation() (src/restorept.c). Route: SRSetRestorePointW via dynamically loaded srclient.dll (this mingw ships no srrestoreptapi.h and no srclient import library — verified — so the documented structs are declared locally and the DLL is loaded at runtime; that is the route that links under mingw, on both architectures). At most one attempt per 24 h (pure decision function, unit-tested across the matrix: local-off, policy-off, policy-on-overrides-local, throttled, boundary-exactly-24 h, allowed); the attempt is stamped before calling Windows. Windows' own throttling is detected by an unchanged sequence number and journaled as "too-recent". Every outcome is journaled: created / too-recent / disabled-by-setting / disabled-by-policy / failed with the Windows status code. Local Settings toggle restore_points (default ON, greyed when the fleet sets it) + fleet policy_restore_points (server: fleet policy field restore_points, unset = device decides — the server half ships in this release without a server version bump, following the 2.5.123 precedent; admin editor under Fleet → policy). Actual point creation verifies on a real Windows box (System Protection tab). - Tests:
tests/client/batch13r3-test.sh — 16 native unit checks (the full restore-point decision matrix; the repair-offer decision) + source guards (protected-folder feed into the rollback module, SHA-256 sidecars + enumerator exclusion, verified-restore pattern, folder restore, 100 MB cap, no second vault module in the build, remediation hook, dynamic-srclient route, policy plumbing both ends, repair-path absence of data-destruction calls, build wiring, What's New guide format).
- Admin: Threat feed manager (Protection → Feed manager). Live per-source counts and one search across everything the web checkers consult —
threats.json domains/IPs, feed_indicators, device-reported iocs, and fleet_iocs bundles (GET /api/admin/feed-manager?q=). Two owner-only, audit-logged actions: remove an entry (POST /api/admin/feed-manager/remove, per source) and allowlist a host (POST /api/admin/feed-manager/allowlist, new checker_allowlist table). The allowlist is consulted first by analyzeLink() — an allowlisted host and its subdomains return Safe even against a condemning feed — making a google.com-style poisoning a ten-second self-serve fix. The UI labels each source's real blast radius: threats.json is served to devices (Web Protection / reputation / web-filter seed), feed_indicators and iocs affect the web checkers (and attribution) only, and fleet_iocs entries were already broadcast to devices at verdict time — removing one stops future checker matches but honestly does not recall copies already delivered. The allowlist likewise changes only the web checkers, never device-side blocking. - Admin: Verdict review queue (Protection → Verdict reports). Works the
verdict_reports table the 2.5.123 report buttons fill (GET /api/admin/verdict-reports?status=, newest first, with per-status counts). Per report, owner-only actions (POST /api/admin/verdict-reports/action, all audited): Allowlist this domain (writes the item-6 allowlist and marks the report reviewed-allowlisted — a test proves a subsequent checker verdict flips to Safe), Verdict was correct (reviewed-correct), and Dismiss (dismissed). - Admin: One-click data backup download (System → Maintenance).
GET /api/admin/backup/download (owner-only) builds and streams a .tar.gz of the data directory with MANIFEST.txt inside listing the included entries and the exclusion rules. Exclusions are enforced twice (entry selection + tar --exclude) and printed in the manifest: session_secret, totp_secret, totp_recovery, vault/ contents, backups/ (no recursion), and any filename suggesting .env/secret/credential/Stripe material; Stripe keys are environment variables, never files, so they cannot be in the archive. Every download writes a backup.download audit entry. This is separate from the scheduled whole-directory backups, which keep their existing rules (and optional encryption). - Admin: Trial funnel (Customers → Trial funnel).
GET /api/admin/trial-funnel returns the last 8 weeks of trial cohorts with the definitions stated on the section: Started = was_trial licenses created that week; Activated = cohort trials with at least one registered device (any time since); Converted = cohort trials that later gained a paid Stripe subscription. Seeded two-week fixtures in the test suite reproduce the counts exactly at each stage. - Admin: Login security watch (System → Login security). A new
login_attempts journal records password-step outcomes (surface customer/reseller/admin, email, IP, success — never passwords or 2FA codes) from /api/login, /api/reseller/login, /api/admin/session, and failed admin Basic auth. GET /api/admin/login-watch shows the last 50 failures, top offending IPs (24h and 7d), and current blocks. Block IP / Unblock (blocked_ips table, owner-only, audited) is enforced by a gate on customer login, the 2FA step, registration, and the public checker endpoints (/api/link-check, /api/scam-check, /api/verdict-report) — and deliberately never on /admin or /api/admin/*, so the owner cannot lock themselves out of the panel that unblocks. Tests prove a blocked IP gets 403 on customer login while admin login from the same IP still works, and that unblocking restores access. All five additions are documented in the admin guide. Server-only release; the Windows client stays 2.5.114.
Version 2.5.154
Spanish round 3 — My Account is fully translated: the route-built license cards (platform and device lists, status badges, the custom-plan block, transfer/remove/delete actions, switch-to-yearly, and the complete 2.5.151 subscription-pause copy — scheduled and in-pause badges and explanations), the referral and redeem cards, family sharing and shared-license cards, and every account-tab renderer (billing, sessions, sign-in history, the email preferences center, threat alerts, API keys, webhooks, fleet report, remote support, network devices, two-factor and passkeys, white-label branding). The download page's "For your device" badge and the subscribe page's yearly-view preset note (2.5.131) are translated too, and the forgot/reset password pages from 2.5.151 were verified already bilingual — so no 2.5.151–153 customer string is born stale in Spanish. Same machinery as rounds 1–2 (?lang=es / aelang cookie, static chrome only; dynamic data identical). Email templates remain English — that is the standing convention. Suite: v25154-batch13r2-test.js.
Version 2.5.153
Batch 13 release A (server-only; the Windows client stays 2.5.123):
- Helper-admin two-factor authentication (item 6). Until now the only admin 2FA was the owner's single global TOTP; helper admins (2.5.137) signed in with a password alone. Each helper now enrolls their OWN authenticator: new
totp_secret / totp_recovery columns on admin_users, enrollment under System → Admin security → Helper admin two-factor (one-time secret + confirm code, reusing the same RFC 6238 primitives as the owner's flow), and 10 single-use recovery codes per helper, scrypt-hashed at rest and shown exactly once. An enrolled helper's sign-in requires the code step (a recovery code works in its place and is burned on use). New owner setting "Require 2FA for helper admins" (GET/POST /api/admin/2fa/settings, owner-only): a helper who has not enrolled signs into a confined session that can reach ONLY the panel shell, whoami, logout, their own password change and the 2FA self-service endpoints — a gate, not a lockout — and the panel opens on a banner that takes them to enrollment. The owner is exempt from the toggle by construction. The helpers list now reports each helper's enrollment state (never the secret). Suite: v25153-batch13r1-test.js covers enroll → code demanded → wrong code refused → right code passes → recovery code works exactly once → toggle forces enrollment → owner unaffected. - Status-incident workflow + timeline (item 7). Incidents gain a
status (investigating → identified → monitoring → resolved) and a source marker (additive columns; everything written so far is manual), plus a status_incident_updates timeline table. POST /api/admin/status-incidents accepts an opening status; new POST /api/admin/status-incidents/update appends a timeline message and/or moves the status — resolving through it stamps resolved_at and sends the same subscriber resolution email as the classic resolve endpoint, and moving a resolved incident back reopens it. The public /status page renders the current status and the full timeline for active incidents and recent history; uptime still computes from started_at/resolved_at exactly as before. Roles: owner + Manager (deliberately NOT added to the owner-only denylist; the Support allowlist does not carry these paths, so Support is refused). - Bulk fleet actions (item 8). New
POST /api/admin/installs/bulk-action + buttons on the Fleet installs bar: Restart (with a Force option → restart-device:force), Isolate and Release across the selected installs, reusing the 2.5.123 command machinery (restart via the JSON command store; isolate/release via queueDeviceCommand + the same EDR isolation event the single-install action writes). Unlike the 1.4.6 bulk-command (counts only), this endpoint reports per-install outcomes and audit-logs the batch. No bulk path in the codebase capped its selection before; this one sets the pattern — at most 50 installs per call. Fleet-scan behavior is unchanged. - Support inbox canned replies (item 9). New
support_templates table + CRUD in the Support inbox section, gated by exactly the access that can open the inbox (the Support allowlist carries the template paths because Support can reply). Each reply box gains a template picker; inserting substitutes {name} from the ticket's own name field (falling back to "there" when the ticket has none, so greetings stay grammatical). - Security-event export (item 10).
GET /api/admin/installs/:id/events/export?format=csv|json (owner + Manager) streams one chronological file merging detections, security_events, tamper_events and device_commands with their ack/result outcomes, named after the install's label. The CSV writer is real RFC 4180 escaping; an unknown install returns a 404 JSON. Linked from the install detail page (Event export card).
Version 2.5.123
Batch 12 release C — the client release, with its server fields in the same commits (the server version stays 2.5.152):
- USB storage control — the wiring is complete (item 6). The fleet policy itself (
usb_storage: allow|block, admin Fleet UI, client USBSTOR Start=4 enforcement with backup/restore, per-minute re-enforcement, SetupDi disable on arrival) dates to 1.26.0/1.27.0; what was missing is now built: the heartbeat returns the boolean form policy_usb_block next to policy_usb_storage and the client applies it through the same usbctl_policy_storage() path the string form uses; the client reports mod_usb_block in the heartbeat module states (new installs.mod_usb column), and the admin installs list + install detail show a USB module badge (green = storage BLOCKED, grey = allowed — red would wrongly read as a fault). Engagement warning: when the block engages with removable drives already mounted, those drives stay visible until replugged (USBSTOR Start=4 only stops *new* devices) — the client now says exactly that: a journal warning, a usb_block_engaged EDR event naming the drive letters, and a tray note, once per block period, instead of pretending the drives were ejected. The unmanaged local tray setting (Settings → USB storage, default allow) is unchanged and shares the same apply path. - Updater architecture fix (item 7). One shared selector,
ae_update_pick_exe() in cloud.c: 64-bit builds take the manifest's exe_x64_url/exe_x64_sha256, 32-bit builds take exe_url/exe_sha256, and there is deliberately NO fallback to the other architecture when this one's entry is missing (the entry is simply unavailable). The offline folder import (2.5.122) used this rule from the start and now calls the shared function; the ONLINE checker previously read exe_url unconditionally — 64-bit clients downloaded the 32-bit installer — and now routes through the same selector. Unit tests pin all four combinations (both arches × present/missing entry) plus the offline check's unchanged refusal codes. - Windows Security Center presence (item 8) — display-level registration, with the audit stated plainly. New module
src/seccenter.c: at service start the client publishes an AntiVirusProduct instance (displayName, a stable instance GUID, paths to its own exe) into the WMI namespace root\SecurityCenter2 via IWbemServices::PutInstance (create-or-update), refreshes it after heartbeats whenever the computed state changed (10-minute backoff on failure), and deletes the instance at uninstall. productState comes from the pure, unit-tested seccenter_product_state(rtOn, sigFresh, snoozed): the healthy state is the value shipping products publish for "on + up to date" (266240 / 0x41000); stale signatures (>72 h) set the out-of-date nibble; protection the user paused reports *snoozed*; and anything not actually protecting — protection off, expired license, or a paused subscription — never reports "on". Every failure is fail-open with a log line; a broken WMI stack can never affect protection. THE AUDIT, so nobody overclaims: this is a *display-level* presence. It makes AEAntivirus visible to WMI consumers that enumerate installed antivirus products. It does NOT make Windows Security treat AEAntivirus as the registered AV provider: Microsoft's Security Center learns a product's state through a private API available only under NDA to members of the Microsoft Virus Initiative (MVI). MVI membership in turn requires a commercially available real-time-protection product, an NDA + program license agreement, an active industry reputation, Trusted Signing for the security solution, and independent certification renewed every 12 months; kernel-level pieces (ELAM drivers) additionally require Microsoft-signed drivers via the Hardware Certification Kit. Until AEAntivirus is an MVI member with signed binaries and current certification, "registered with Windows Security" / "integrated with Windows Security" claims would be false, and no product copy makes them. (The WMI write itself also cannot be exercised from the Linux build VM — it is build-verified and unit-tested at the state-computation layer, and wants a on-PC confirmation after install.) - Restart device command (item 9). New fleet command in exactly two forms —
restart-device and restart-device:force — validated identically at every layer: server VALID_CMDS, remoteCommandIsValid (any other argument rejected), the client command parser cmd_parse_one, the heartbeat-channel validator rcmd_valid, and both dispatchers. Admin install detail gains Restart device / Force restart actions (confirm dialogs; the queue endpoint's audit entry covers them). Execution (src/restartctl.c): the countdown dialog lives in the tray — "Your administrator scheduled a restart", a live 60-second countdown, a Cancel restart button in the normal form; the force form has no Cancel and ignores the close box. It is a singleton twice over (named mutex Local\AEAntivirus_RestartDlg across the tray/service process pair + in-process guard) and follows the modal discipline (its loop exits via its open flag; no PostQuitMessage from WM_DESTROY). If the service polls the command first (session 0, no desktop), the request is persisted to restart_pending.txt and the tray picks it up — the announce.c routing — and the service does not ack. Outcome reporting (minimal extension): the ack now carries an optional result — on expiry the client acks restarting FIRST, then restarts via InitiateSystemShutdownEx (shutdown privilege enabled, planned application-maintenance reason, ExitWindowsEx fallback); on cancel it acks canceled and journals the cancellation. Server-side, SQL commands store the result on the new device_commands.result column; JSON-store commands (the admin queue) record it on the queue entry, matched by command text, and the install detail's command history shows it. - Tamper-event reporting (item 10). Verified first: client journal entries already sync to the server's
security_events timeline via POST /api/security-events. The missing span is built: client call sites journal with type tamper through the new ae_tamper_event() helper — (1) three consecutive tamper-PIN failures in the PIN dialog (streak resets on success), (2) a protected-settings save refused for a missing/wrong current PIN, (3) a service STOP/SHUTDOWN control while a tamper PIN is armed. Honest scope on (3): the service control handler still *accepts* stops (the SCM owns the lifecycle; real stop-resistance is the service's failure-action restart + watchdog) — what ships is the *report*. Server-side, the security-events ingest mirrors genuinely-new tamper rows into a dedicated tamper_events table (install, kind, detail, at) — chosen over reusing security_events alone because the admin views need a per-install recent list and a 7-day count without scanning the general timeline. Tamper events are NOT detections: ingest never touches the threats_stopped counter, the install's lifetime counter, or the detections table (the suite proves all three unchanged). Admin: the install detail shows the last 10 tamper events in a Tamper events section, and the installs list badges installs with tamper events in the last 7 days. - Paused-subscription client wiring (completes item 2). Server 2.5.151 already returned the distinct
paused license state; client 2.5.123 now handles it as its own state: protection is OFF for the whole pause window with no 72-hour grace (a pause is deliberate, not a billing failure), update checks are skipped like an expired license, the payment prompt never fires for it, and the tray banner reads "Subscription paused until <date> — protection is off during the pause" with the payment button hidden. Protection returns automatically on the first heartbeat after the subscription resumes. (The previous client, 2.5.122, treats the unknown paused value as a generic bad state: 72-hour grace, then off, with no paused label.) - Also in this push (server, no version bump):
GET /api/admin/installs fixed — paused_until was ambiguous between the installs and licenses tables since licenses gained that column in 2.5.151, and the endpoint has returned 500 since; the column is now qualified. Caught by this release's server suite. - Tests:
v25123-batch12r3-test.js (server: USB policy round-trip + module state, restart validation/delivery/outcome on both command stores, tamper ingest with counter-invariance proofs, /admin inline-script regression) and tests/client/batch12r3-test.sh (selector + productState + parser unit tests, source guards for all five items, guide format). What's New in the client covers each change in one line each.
- "Report a wrong verdict" on both free checkers. The result view on
/link-checker and /scam-checker now carries a Report a wrong verdict button with an optional note (POST /api/verdict-report, no login, per-IP throttled 30/hour). Reports land in the new verdict_reports table (tool, subject, verdict, reviewer detail, visitor note, status new) for the admin review queue shipping in 2.5.124. Privacy: for the scam checker the raw pasted message is *never* stored — the endpoint re-runs the analyzer in memory and persists only the verdict, the flag labels, and the hosts of links extracted from the message (the checker's no-storage promise now covers reports too). For the link checker the submitted URL is the report subject and is stored with its parsed host. If the visitor's claimed verdict differs from a fresh analysis, both are recorded for the reviewer. - "AEAntivirus for Business" page (
/business). A dedicated pitch for the Business and Custom tiers: plan cards, the Custom volume-discount ladder (10% off at 11+ devices, 20% at 25+, 30% at 50+), and the fleet-management story (centralized policy, live fleet view, remote actions, license control, audit trail), with CTAs to /subscribe and /contact. All pricing is computed at request time from the subscribe page's own sources (TIERS, CUSTOM_VOLUME_TIERS, customEffectivePerDeviceCents()) — the view holds no price numbers of its own, so the two pages cannot drift apart. - Free tools hub (
/tools). One shareable page collecting the link checker, the scam message checker, and the Threat Guide, linked from the footer (Support column) and the sitemap. - Customer reviews — homepage section + admin manager. The homepage gains a "What customers say" section that renders only owner-approved rows from the new
reviews table, and is omitted entirely — no placeholder, no invented quotes — while zero reviews are approved. The owner manages them in the admin panel (Customers → Homepage reviews): add (name, optional town/state, star rating, text), approve/unapprove, delete (GET|POST /api/admin/reviews, POST /api/admin/reviews/moderate, POST /api/admin/reviews/delete; mutations owner-only, all audited, output escaped). The feature ships empty: it appears the first time Austin enters and approves a real customer review. - Customer 2FA — already shipped in 1.7.0; now covered by an end-to-end regression suite. Per-account TOTP (
/account security section, RFC 6238 via node:crypto, reusing the shared base32Decode/totpCode/verifyTotpCode helpers the admin and reseller flows use), one-time scrypt-hashed recovery codes, a need_2fa second step at POST /api/login + POST /api/login/2fa, 5-attempt/15-minute lockout, and password + current-code required to disable. This release's v25123-site3-test.js drives the full flow over HTTP (enable → login with code → wrong code rejected → recovery code works exactly once → disable restores plain login → accounts without 2FA unaffected). One deliberate deviation from a plain reading of the batch-3 spec: disable requires both password and a current code (the shipped 1.7.0 behavior) rather than either one — a hijacked session alone, or a stolen password alone, cannot strip the second factor. Help/FAQ gains a 2FA walkthrough and the Terms' Account security section now states that recovery codes are the customer's to keep (shown once, stored hashed, support cannot re-issue them). Server-only release; the Windows client stays 2.5.114.
Version 2.5.152
Batch 12 release B (server-only; the Windows client stays 2.5.122):
- Notify customer from a detection (admin Detections board). Each detection row gains a Notify customer action: a compose view prefilled from the new
detection_notice email template (editable in Customers → Email templates; variables device, threat, action, time — device label, threat name, what was done, detection time) plus an optional admin note appended to the send. POST /api/admin/detections/:id/notify (owner-only) sends via sendMail to the owner of the device's license (install → license → account email, falling back to the license's email), stamps the detection with new notified_at / notified_by columns, appends "Customer notified <time>" to the triage note through the 2.5.148 triageSet store (state preserved), audit-logs the send, and the send lands in mail_log (kind detection_notice). If the install's license has no reachable owner email, the list API reports the reason and the UI disables the button with it; a second notify is allowed and the compose view warns "Already notified at <time>". A failed send records nothing. Suite: v25152-batch12r2-test.js proves the rendered variables in the captured mail, the marker + triage note, the no-email refusal, and that an admin-edited template override is what actually gets sent. - Device exclusions viewer (admin install detail). Mechanism map: per-device false-positive exclusions live in the
fp_exclusions table (install_id, path, threat, created_at — written since 2.5.82/2.5.91 when a detection is marked FP, by the admin verdict flow or the client popup); devices pull them through GET /api/fp-exclusions (install-authenticated) into the client's local store, and the engine never flags those paths again on that device. The install detail gains an Exclusions section (GET /api/admin/installs/:id/exclusions) listing that install's rows — path, threat, marked-at, and the originating detection (latest detection with the same install + path) where known — each with a Remove action (POST /api/admin/installs/:id/exclusions/remove, owner-only, confirm in the UI, audit-logged) that deletes the enforcement-store row, so the device stops honoring it on its next sync. The suite asserts scoping (another install's rows untouched), propagation (the device-facing sync no longer returns a removed path), and that no customer-facing surface exposes exclusions.
Version 2.5.122
Batch 11 release C — the client release, with its server policy fields in Server 2.5.149 (above):
- App allowlist (strict) mode (item 6). Fleet policy gains
appfw_mode (blocklist default | allowlist) and policy_allowed_apps, plumbed like blocked_apps end to end (admin Fleet UI text areas + mode/metered selects, sanitizePolicy, heartbeat delivery). Client: cloud.c applies both (stash/restore into config allowed_apps), firewall.c gains fw_app_is_allowed() / fw_app_allowlist_mode(), and the fwapp poll gains an allowlist branch ahead of the prompt flow: approved = remembered user allow decision or the policy list (Windows system programs still pass silently, the admin blocklist still wins in both modes); anything else is blocked immediately via the per-app netsh rule and asked about through the existing prompt (Yes = always allow → remembered allow rule; No = keep blocked). On policy-locked (managed) devices there is no prompt — the block stands. Default mode is unchanged. - "Why was this flagged?" (item 7). The heuristic scorer now retains its top contributing signals per detection: a static plain-English map (one sentence per signal tag, in
src/aiscore.c) plus ai_top_reasons() rank the signals that actually fired for that file by active weight (positive contributors only — a score-lowering signal never explains a flag). Every detection stage in scan_file_desc records its real reason (signature hits name the signature; cloud verdicts, fuzzy matches, cert blocklist, custom rules, YARA, fleet intel, and classic heuristics each record their own). The strings ride the journal detail, the detection popup ("Why this was flagged:" in the verdict dialog), and the server report as reason_texts — a NEW field, deliberately separate from the ML signal-tags reasons column that verdict learning consumes. Server 2.5.149 stores them on a new detections.reason_text column (JSON array, max 3, string-only, ≤200 chars) and the admin Detections list shows them as a Why: line (parsed why[] in the API). - Protection score (item 8). Status page shows a 0–100 score computed only from real local state via the existing getters: protection 60 (real-time 20, firewall 15, web protection 15, advanced RTP 10; off or paused = 0), signature freshness 15 (≤24h full, ≤72h half — from signatures.db mtime, falling back to the client binary's own mtime for the builtin set), scan recency 15 (≤7d full, ≤30d half, from the stats file's last_quick/last_full), up to date 10 (from the
update_pending marker the update checker now maintains). The pure function protection_score_compute is unit-tested across the matrix (each degradation costs exactly its weight). The score is additive only — the amber "Attention needed" logic is untouched — and the formula is printed next to the score in the UI and in What's New. - Metered-connection mode (item 9) — BUILT, not blocked. Verified in this environment: mingw ships
netlistmgr.h and an INetworkCostManager probe compiled and linked for BOTH x86_64 and i686 with the already-linked -lole32 -luuid (vtable calls, no COBJMACROS dependency). The client queries GetCost and classifies with the pure ae_cost_is_metered() (VARIABLE / OVERDATALIMIT / APPROACHINGDATALIMIT / ROAMING, or FIXED without UNRESTRICTED). When metered and respected, cloud_check_updates still fetches the (tiny) manifest but defers all signature-pack downloads, the ML model, and the engine download; update_pending/update_deferred_metered are set and the tray shows "Updates are waiting for an unmetered connection." Heartbeats, scans and protection never defer. Local toggle in Settings (default ON); fleet policy policy_metered_mode (local/respect/ignore) forces the behavior, and the Settings box greys with an explanation when forced. Unknown cost never defers. - Offline update import (item 10) — with one honest correction to the spec's premise. There is NO manifest signature or baked-in key anywhere in the online channel: the online updater's verification is TLS origin + the exe's SHA-256 matching the manifest + version-newer (verified in
cloud_check_updates, build.sh, and publish-update.sh; the manifest carries no signature field). The import therefore reuses exactly that chain — no new key infrastructure was invented, and the spec's "wrong-key manifest" test maps to the tampered-exe hash refusal and malformed/missing-field manifest refusal. The staging block was factored into cloud_stage_engine_update() shared by both paths; the pure verifier ae_offline_bundle_check() picks the architecture entry (_WIN64 → exe_x64_*), requires version + entry + hash, and demands version-newer. UI: "Import update from folder…" beside Check for Updates; any failure produces a standard message box naming the reason and installs nothing. The /download page (Server 2.5.149) now explains saving manifest.json + the installer for offline use. (Noted in passing: the ONLINE checker has always used the x86 exe_url entry for both architectures; the import picks the arch-correct entry. Changing online behavior was out of scope.) - Docs: homepage rotates (Client 2.5.122 + Server 2.5.149 visible; 2.5.147/2.5.146 fold), Help & FAQ gains five entries (why flagged, protection score, metered connections, offline updates, allowlist mode), the admin guide gains the 2.5.122 fleet-fields entry, the client What's New gains the 2.5.122 entry (CRLF, one line per change). No ToS change: no described term changed.
- Suites:
v25122-batch11r3-test.js (server: policy sanitize/persist/heartbeat round-trip, reason_texts sanitize/store/admin round-trip, download copy, manifest at 2.5.122) and tests/client/batch11r3-test.c/.sh (reasons from the real breakdown path incl. the trigger feature shape and the negative-weight honesty case; score matrix; metered flag matrix + deferral decision; offline bundle accept/refuse matrix; source guards for the Win32 wiring).
- Fix: the link checker condemned google.com (production bug, found by Austin).
POST /api/link-check returned Dangerous/100 — "on our threat-intelligence blocklist" — for google.com, www.google.com, and https://www.google.com. Root cause: the feed pipelines reduce malicious URLs to bare hosts at ingest (hostOfUrl() discards the path) and merge them into the same domain lists the checker reads — and real feeds (URLhaus/OpenPhish/PhishTank-style) list URLs hosted on major domains, including google.com/url?q=… open-redirect links and Docs phishing forms. A one-page report thus became a whole-domain condemnation of google.com in threats.json, which linkIocHit() checks first. The 2.5.118 tests missed it because the seeded test data contained no real feed entries. - Precise indicator matching (
splitIndicator() / indicatorHitLevel()), applied to every checker source — threats.json domains, feed_indicators, client-reported iocs, and fleet_iocs (whose URL check previously condemned a bare host for ANY fleet URL on it, with no path comparison at all): (1) a bare-host indicator condemns its own host and subdomains of it — matching only ever walks down from the indicator, never up to a parent or across to siblings (a listed docs.google.com no longer touches google.com or mail.google.com); (2) a path-carrying indicator condemns only submissions on the exact host whose path starts with the indicator's path — the actual reported phishing form still flags Dangerous, the rest of the host does not; (3) major-domain guard: for hosts on the well-known list, a bare-host feed hit alone caps at Caution with an honest reason ("specific pages under this domain appear in threat reports… this exact link is not one of them"), and a Caution-grade hit also caps the final verdict so stacked soft signals can't re-condemn the domain. Exact host+path hits still condemn a major domain outright. - The scam checker (2.5.120) is fixed by the same change — it scores embedded links through the shared
analyzeLink(), so a pasted message containing a google.com link is no longer condemned by feed entries for unrelated Google pages. - Regression coverage (
v25122-linkcheck-precision-test.js, two seeded boots): feed indicators seeded with a docs.google.com form URL and a google.com redirect-path URL — google.com / www.google.com → Safe, mail.google.com → not Dangerous, the exact form URL → Dangerous, a sibling Docs path → untouched. Canary sweep (google, microsoft, apple, amazon, facebook, youtube, github) with poisoned subdomain + redirect feed entries: no canary root returns Dangerous. A bare google.com entry in threats.json (the production shape) caps all Google hosts at Caution, never Dangerous, while the exact form URL stays Dangerous. Subdomains of genuinely listed bad domains still flag Dangerous. Server-only release; the Windows client stays 2.5.114.
Version 2.5.121
- Reset statistics button (Performance page). Puts this PC's threats-stopped and lifetime counters back to zero and clears the local detection history (the Security Hub timeline journal + the YARA findings store), after a standard confirm that states the scope plainly. The reset runs identically for the button and for the server's new
reset-stats command — one mechanism, three doors. Because the tray and the service are separate processes, the button also drops a flag file the engine loop consumes, so the service applies the reset to its own counters too and can never write the old totals back from memory. Quarantine contents and false-positive exclusions are deliberately preserved; a single journal entry records that the reset happened. New headless suite (reset-stats-test.sh, 26 checks) drives the exact shipped reset code against temp files: both counter keys zeroed, every other stats key preserved, journal deleted, unrelated stores untouched. reset-stats fleet command. Added in all four places (server VALID_CMDS, server remoteCommandIsValid, client cmd_parse_one exact-match whitelist, engine dispatchers): the command poll executes it through engine_run_command and acks only after the reset has run, which is what lets the server lift its heartbeat guard at exactly the right moment. cmd-test gains the parse cases (bare accepted; argument and near-miss forms rejected).
- Admin: Win-back list (Customers → Win-back). Licenses with status canceled/expired/deleted whose end falls within the last 30 days, most recent first (
GET /api/admin/winback). "When it ended" uses the timestamps the schema actually has: deleted_at for customer-deleted licenses, otherwise current_period_end — there is no status-change timestamp, and the section says so. Per-row Send win-back email (POST /api/admin/winback/send, owner-only) sends a friendly come-back message linking /subscribe and /account (checkout with the same email reactivates immediately via the existing payment webhook), stamps the new licenses.winback_sent_at column (re-send allowed), and writes a winback.email audit entry. SMTP-unconfigured returns the same clear visible 400 as the 2.5.119 expiring-soon reminders — never a silent failure. - Admin: Stripe promo code manager (Customers → Promo codes (Stripe)). Stripe-native codes, distinct from the existing local Promo coupons section (which this server validates itself): create a percent- or fixed-amount-off code with optional max redemptions/expiry (
POST /api/admin/promo-codes, owner-only — creates the Stripe coupon + promotion code via the API), list codes with status and redemption counts (GET /api/admin/promo-codes, from Stripe with a 60-second in-memory cache, source labeled in the UI), and deactivate (POST /api/admin/promo-codes/deactivate, owner-only). So the codes actually redeem, checkout sessions now set allow_promotion_codes: true whenever no server-side discount (promo/referral/volume) is applied — Stripe forbids combining the two, so server discounts keep precedence. - Admin: Block fleet-wide from any detection. Detections rows gain a Block fleet-wide button (
POST /api/admin/detections/:id/block-fleet, owner-only, confirm in the UI). It writes through the exact paths confirmed-threat verdicts already use — no parallel store: the file hash is upserted into community_threats (with fleet correlation), and domains/URLs/IPs from extractIocs() are pushed via broadcastIocs() into fleet_iocs + queued device commands. The row is stamped with the new detections.fleet_blocked_at column and renders a Blocked fleet-wide badge; re-clicking is an idempotent no-op. A detection with no valid hash and no network indicators gets a clear 400 explaining why. Every block writes a detection.block_fleet audit entry. Verified end-to-end in tests: after a block, the hash is served by the client-facing GET /api/community-threats/recent feed. - Admin: customer notes. Admin-only notes per customer account in the account detail view (Customers → Accounts → click an account): add/list/delete via
GET|POST|DELETE /api/admin/accounts/:id/notes[/:noteId] (mutations owner-only), stored in the new admin_notes table, timestamped, escaped on render. Adds/deletes are audit-logged without the note content, and a test asserts the note text never appears on the customer's /account page. - Admin: Fleet pulse charts (Dashboard → Fleet pulse).
GET /api/admin/fleet-pulse (60-second server-side cache) aggregates the detections tables: detections per day for the last 30 days (zero-filled), top 8 threat types, and the 8 riskiest devices by detection count (hostname joined from installs). The section renders self-contained inline SVG charts — no external libraries — under the 2.5.112 independent-loader pattern. All five additions are documented in the admin guide (Help tab). Server-only release; the Windows client stays 2.5.114.
Version 2.5.120
- Explorer right-click scan, now in the app. The "Scan with AEAntivirus" verb has been registered on files and folders since 1.4.4 and removed on uninstall — but it ran a headless console scan. The verb now launches
--scan-target "%1": the target is handed to the tray UI through a pending file in the data dir (the same pattern as announcement handoffs), so the app opens on the Status page and runs the targeted scan where you can watch it (engine_scan_path, logged as a "custom" scan, same one-scan-at-a-time claim as Quick/Full). If the app is already running, the launcher detects the tray mutex, leaves the pending file, and exits — the running tray consumes it on its 3-second command poll (deleting the file *before* the scan starts, so a crash can never re-fire it). With no tray running, the launcher process becomes the tray and consumes the file at startup. The old --scan <path> console scan remains for scripts. (Verified here by source guards + the headless scan machinery it reuses; the click itself is verified on a real PC.) - Quarantine auto-clean. Quarantined items older than an age limit are now permanently removed through the exact code path of the manual Delete button (file + record, across both modes' quarantine DBs — the 2.5.117 dual-dir lookup) at startup and once a day, with each batch logged. Effective age resolution: a fleet policy value wins (
quar_autoclean_days, sanitized 0..3650 server-side, delivered in the heartbeat as policy_quar_autoclean_days with a control in Protection → Policy → Updates & patching), else the local setting (default 90 days, cycled 7/30/90/never on the Power User page, whose "Clean now" button now runs this record-safe clean instead of the old file-mtime sweep that could strand records); 0 = never. The quarantine window shows the last run ("Auto-clean: N old item(s) removed on <date>"). Schema finding, stated plainly: quarantine records (id|timestamp|origpath|threat|desc) carry no false-positive-candidate state, so age is the only rule — and it is safe that it is, because a file the user marks as a false positive is restored *out* of quarantine by the FP flow and is never sitting there to be cleaned. Age math is whole days from the record timestamp, unit-tested at the 90-day boundary. - Installer styling pass, closed out. Audit result: the setup wizard, the 2FA prompt dialog, and the Terms/License viewer already adopted the shared app theme (dark background, cyan accents, owner-drawn buttons) in 1.97.0, and the Linux installer gained its AE banner + NO_COLOR/TTY-aware colored output in the same wave. The remaining gaps are closed here: the doc viewer's Accept/Decline buttons now receive the app font (the owner-draw painter uses the DC font — they were silently falling back to the system default) and the viewer class gets the standard arrow cursor. Legal text stays UI_TEXT on the panel background (high contrast). Visual verdict on a real PC is still the final check, as always.
- Build hygiene.
build.sh now stamps the three .rc files' version metadata (FILEVERSION/PRODUCTVERSION + FileVersion/ProductVersion strings) from APP_VERSION in src/cloud.h at build start — the checked-in .rc files remain the template, and after any build they match APP_VERSION exactly; the hand-sync step that used to ride along with every client release is gone. A new guard (tests/client/rc-version-guard-test.sh) fails when they disagree. Also removed from the repo: the stale linux/aeantivirus-linux.tar.gz duplicate (53,280 bytes, pre-1.23.1) — the canonical Linux tarball lives in server/installer/ + server/updates/ (83,163 bytes, sha256 295371f3…, byte-identical) and is untouched. Linux stays 1.23.1: its installer styling already shipped in 1.97.0, so no tarball change and no version tick. - Server note: this release's only server change is the
quar_autoclean_days fleet-policy field above — shipped under the existing Server 2.5.142 version (no bump), with its own suite (v25142-batch10r3-test.js: defaults, save sanitizing, heartbeat delivery).
- Scam message checker (
/scam-checker + POST /api/scam-check). Anyone can paste a suspicious text or email and get a verdict — Likely scam / Suspicious / Looks okay — with the specific red flags found: gift-card and crypto payment demands, wire/payment-app pushes, password/PIN/one-time-code requests, urgency and threats, bank/IRS/brand impersonation, secrecy demands, prize claims, remote-access asks, family-emergency patterns, and too-good-to-be-true money. Any links inside the message are scored with the exact same analyzeLink() engine as the 2.5.118 link checker (shared code, verdict Dangerous +45 / Caution +15). The family-emergency signal requires distress context — an ordinary message to grandma scores nothing (a bare kinship word used to false-flag it; caught by this release's tests). Rate-limited to 30 checks/hour/IP. Privacy: the pasted text is analyzed in memory only and never stored, logged, or written to the database — the page says so, and a test pins the behavior. - Referral program — "Give a month, get a month". Every account gets a personal referral code (
referral_codes) and link on My Account, with friends-joined and free-months-earned stats. A visitor arriving via /?ref=CODE gets a 30-day aeref cookie; at checkout (when no promo code is used) the server-managed Stripe coupon aeav-referral-free-month (100% off, once, created via the API like the 2.5.116 volume coupons) makes their first month free, and a referrals row is recorded at checkout.session.completed. When that subscription's first invoice.payment_succeeded arrives, settleReferralReward() applies the same free month to the referrer's active subscription — exactly once (only pending_payment rows transition, so duplicate webhooks can't double-credit; Stripe discount state is re-checked so a second free invoice can't stack either). Guards: self-referral refused by account or email, one referral per referred account/email, cap of 3 earned free months per calendar month, and if the referrer has no active paid sub (or it already carries a discount) the reward is held with the reason recorded instead of clobbering anything. - Gift subscriptions. The subscribe page gains a "This is a gift" toggle + recipient email. Checkout carries
gift/gift_email metadata; the webhook provisions the license for the recipient — email = recipient, account_id NULL (new licenses.gift_email / gift_emailed columns), so it never appears among the buyer's licenses and the buyer's trial installs never convert onto it. The recipient is emailed the key + setup instructions via the existing mailer; if SMTP is unconfigured, GET /api/license-for-session reports gift/gift_emailed and the success page shows the buyer the key with clear "send this key to them yourself" messaging — no silent loss. Recipients claim the key with the new POST /api/account/redeem-key on My Account (unknown/deleted key 404, key claimed by another account 403, own key a no-op success) — or automatically when they register with the gift email, via the existing email-match linking. - Scam of the Week (homepage). A spotlight card rotates weekly through the Threat Guide's scam entries, picked deterministically from the ISO week number (same week → same scam everywhere, no stored state). The guide's data array was lifted to module scope (
GUIDE_GROUPS) so the spotlight derives from the exact source the guide renders; the card links to the entry's /threat-definitions#anchor and cross-links the scam checker. - Switch to yearly in My Account (
POST /api/account/switch-yearly). Fixed-tier monthly subscribers see a "Switch to yearly — save X%" button on the license card, shown only when a yearly Stripe price is configured for their tier (STRIPE_PRICE_ID_<TIER>_YEARLY) and Stripe confirms the subscription is currently on the monthly price; the save % is computed from the live Stripe prices (cached 1h). The endpoint re-verifies ownership and eligibility, swaps the subscription item to the yearly price with proration_behavior: 'create_prorations', and audits the switch; the existing customer.subscription.updated webhook keeps local state in step. The Custom plan is excluded (per-device monthly, no yearly price) with an explicit 400. Configuration note: the button appears per tier only where Austin sets that tier's STRIPE_PRICE_ID_<TIER>_YEARLY on the host — everything else works with no new configuration. - Docs kept current: Help/FAQ gains gift, referral, scam-checker, and yearly-switching entries; the Terms of Service gain §11 (referral program mechanics) and §12 (gift subscriptions) with the Last updated date moved to October 15, 2026. Footer + sitemap link the scam checker. Server-only release; the Windows client stays 2.5.114.
Version 1.25.1
- Android: brand theme pinned dark. The app theme was
DayNight, which on a phone set to light mode rendered light system-styled widgets and dialogs against the app's forced dark window background — a mixed, off-brand look on exactly the screens this pass covers (first-run/onboarding and the legal dialog). The theme now pins the dark AppCompat variant with the existing brand palette (brand #00E5FF cyan on page_bg #070B12), and the onboarding layout carries the brand background explicitly. No behavior changes. Rebuilt with the proven hand-rolled pipeline and fully re-verified: aapt2 badging shows versionName 1.25.1 / versionCode 27, apksigner verifies, and the dexdump R-class check finds 0 missing library R classes (the check whose absence once crashed the app on a real phone). Still signed with the project's debug keystore, like 1.25.0 — a real release keystore remains Austin's decision before any customer distribution, and the APK has still never run on a device; installing it on a phone is the deciding test.
Version 2.5.119
- Duplicate tray icon fixed (root cause). On a service-mode install (the standard install), two per-session processes each added a tray icon under different single-instance mutexes: the
--notify companion autostarts at logon (HKLM Run + started by the installer), and opening the app started the full tray UI (--tray), which added a second icon with the same uID and tooltip. Nothing stopped both from running in one session, and both persisted because closing the window hides to tray. The fix coordinates the handoff: the notifier now exits at startup when the tray UI already owns the session (checked via the tray mutex before it takes its own), the tray UI stops a running notifier before adding its icon, closing the tray UI in service mode relaunches the notifier (spawned after the tray mutex is released, so its yield check cannot misfire), and tray_icon_add/tray_icon_remove carry an added-flag so one process can never NIM_ADD twice. Source-guard tests pin all four mechanisms; there is no TaskbarCreated re-add path to go stale (explorer restarts were never a duplicate source). Honest limit: the sandbox cannot reproduce a Windows tray — install on a real PC and count icons is the final verdict. - AI Skills Checker now covers the detection engine (Austin's 2026-10-03 request). The existing checker (tray menu → AI Skills Checker) verified the paid AI Assistant add-on against the server; it now also lists this PC's REAL detection components as skills, each with live status read from the subsystem itself: signature database (
engine_sig_count() + definitions set version), heuristic signal scorer (enabled + live flag threshold), trained ML model (mlinfer_loaded() / model version — "Not loaded — heuristic scoring only" when absent, never painted ready), behavior monitors (miner/stalkerware/scareware, reported from real engine/license/pause state), ransomware canary (config state), real-time protection, web protection, and app firewall (their own *_enabled() getters). A Run self-test button scores built-in samples constructed in memory through the real pipeline: a benign text sample (must score below the flag threshold), a synthetic heuristic-trigger PE (deterministic xorshift-filled PE with entry-in-last-section, a writable+executable section, and three non-standard section names, written under a dual-extension name in a temp folder — designed features sum to 100 with built-in weights), and the standard EICAR test string, which is decoded and SHA-256-checked in memory only (it is stored XOR-obfuscated, the 2.5.94 technique, and never written to disk, so the self-test cannot trip the PC's own real-time scanner or another AV). Samples are deleted after scoring; results stay on the device. Sample builders live in the aiscore portable block (aiscore_selftest_*) and are covered by a new headless suite (skills-selftest: benign scores 0, EICAR hash matches the built-in signature value, trigger scores exactly 100 deterministically) plus source guards. Client-only release; the server stays 2.5.140.
- Admin: revenue snapshot on the Dashboard. A new card on the admin Overview shows MRR, active paid licenses, active trials, trials expiring within 7 days, and paid signups this month — computed entirely from local license rows (
GET /api/admin/revenue-snapshot, no Stripe API call): fixed tiers at TIER_PRICE_CENTS list rates, Custom at its effective per-device rate (volume discount included) × device count. "Active paid" = status active with a Stripe subscription on file and not soft-deleted. The local schema stores no billing interval, so yearly plans count at their monthly list rate and the card says exactly that; the Customers → Revenue page keeps the live Stripe view. - Admin: global search. A search box at the top of the console (
GET /api/admin/search?q=) returns grouped matches — accounts (email/name), license keys (key/email), devices (hostname/display name/install id), detections (threat/path) — capped at 8 per group with LIKE fragments escaped. Clicking a result jumps to the section that manages it. Admin-auth only; all output is escaped on render. - Admin: Expiring soon page (Customers). A flat 14-day list (paid + trial + past-due, soonest first) at
GET /api/admin/expiring-soon, separate from the settings-window digest on the Dashboard. Each row has a Send reminder button (POST /api/admin/expiring-soon/remind, owner-only) that emails the customer in the trial-nurture/expiry-warning style (trial / renews-soon / past-due variants), stamps the new licenses.reminder_sent_at column so the row shows when the last reminder went out (re-send allowed), and writes an expiry.reminder audit entry. If SMTP is not configured the endpoint returns a clear 400 and the UI shows it — no silent failures. The list endpoint also reports mail_configured so the page warns up front. - Admin: audit log viewer reads by tail. The Audit log section and its CSV export already existed (filters: text, action prefix, date); what changed is the read path — a new
auditReadTail() reads the log backwards from EOF in 256 KB chunks (buffers concatenated before decoding, so chunk seams can't corrupt a line) and parses only up to the requested entries, instead of auditRead() slurping the whole file on every request. - Admin: Server health page (System).
GET /api/admin/server-health + a new section showing measured facts only: server version, process uptime and start time, memory (RSS, heap used/total), database file size, data-directory total, Node version, and headline counts (installs, accounts, licenses, detections). All new admin sections follow the 2.5.112 independent-loader pattern (own try/catch, visible error + Retry, loaded-mark cleared on failure so the next visit retries). Server-only release; the Windows client stays 2.5.114.
Version 1.25.0
- Android 1.25.0 - first verified build (Android). The Android app's source (current through 1.24.0: account sign-in with 2FA, license registration and trials, Downloads/folder/installed-app scanning, quarantine, attended remote support, QR safety scanner, fleet blocklist and APK scoring) had never been built in this environment - there was no SDK here, and the APK previously published under the Android update feed was a stale 1.23.0-era binary whose feed label did not even match the binary's own version. 1.25.0 changes no features; it makes the build real. The Android SDK (cmdline-tools 23.0, platform-34, build-tools 35.0.1) is now installed persistently at
~/workspace/android-sdk, and the app builds end to end with the project's own manual pipeline (android/build-tools-local/build-manual.sh: AAR extraction, aapt2 compile/link, generated library R classes via gen_lib_r.py, javac + kotlinc, d8, zipalign, apksigner) with a single pre-existing deprecation warning in the QR scanner. Verified on the shipped APK (3,679,844 bytes, SHA-256 c7e168a5843d6998747a5896afcb3a126ec1a98a6887eeee337c851ec643520d, versionCode 26 / versionName 1.25.0): apksigner verify passes; badging confirms package net.aeantivirus.android with launcher activity MainActivity; and the dexdump R-class check - the check whose absence once let a build reach a real phone and crash with NoClassDefFoundError - passes with 155 R-type references, 155 defined, 0 missing across 39 R packages (the app plus all 38 generated library R classes). The APK is published at server/updates/aeantivirus.apk and server/installer/aeantivirus.apk, and server/updates/manifest-android.json now reads 1.25.0 with the new hash (signature DB v15 and certs v4 fields untouched). Stated plainly: the build is debug-signed with the project's persistent debug key (the same signing identity as prior builds, so it installs over them) - a release keystore is Austin's to provide/decide before any customer distribution - and it has still never been run on a real phone or an emulator (this environment has no KVM), so on-device testing remains an open step. Nothing for Austin to configure on the server.
Network monitor everywhere (server; Windows client 1.24.0; Linux daemon 1.19.0; Android unchanged)
- Windows Network Monitor window — new tray-menu "Network Monitor..." window with a live per-process connection table (process, PID, protocol, local/remote address, state, up/down rates and totals), sortable columns, 2-second auto-refresh, and one-click IP blocking through the existing firewall module (with an under-maintenance notice when the firewall feature is off).
- Linux daemon connection tracking — parses
/proc/net/tcp{,6} / /proc/net/udp{,6}, maps sockets to PIDs, reads per-connection byte counters from nf_conntrack when available (fail-open without it), and reports a summarized snapshot every 60 seconds. - Admin fleet network view — new "Network monitor" section under the Fleet tab: devices reporting, total throughput, per-device top talkers, recent connections, and suspicious destinations flagged against the threat-intel IP list.
POST /api/netstats — install-authenticated endpoint accepting the summarized snapshot (processes ≤ 50, connections ≤ 100, all fields length-capped and sanitized). Server-side 45-second throttle per device; one row per install; rows older than 7 days pruned.- Privacy — snapshots contain connection metadata only (process names, IPs, ports, states, byte counters); never packet contents, URLs, or message text. The Terms now disclose the reporting and the 7-day retention.
- Covered by tests —
server/v1250-test.js (auth, throttling, sanitization, threat-intel flagging, admin section, homepage/terms copy); inline-script regression and the 1.24.0/1.23.0 suites still green.
Version 2.5.118
- What's New guide cleanup (Windows). Austin's screenshot of his 2.5.117 PC showed the in-app What's New / Feature Guide dialog as one unbroken blob: version sections ran straight into each other ("...firing lateWHAT'S NEW IN 2.5.116..."), individual changes were separated by " - " inside a single long paragraph, and hard-wrap continuation indents rendered as double spaces. Root cause: the dialog is a standard multiline EDIT control filled with
SetWindowTextA, and those controls only break lines on CRLF - the What's New entries written since 2.5.93 used bare \n (LF), which the control silently swallows, so every heading, bullet and wrapped fragment collapsed into one paragraph. The whole visible guide text (all 92 What's New sections from 2.5.118 back to 1.84.10, the Feature Guide body, and the pinned 2.5.67 note shown above it) is now written in CRLF with one format throughout: each WHAT'S NEW IN X.Y.Z heading on its own line, each change on its own single line starting - , rewritten short and plain (one idea per line - the multi-clause explanations were cut down to what changed and why it matters), a blank line between sections, no double spaces, no dash-glued words, no mid-line section starts, ASCII only (the stray em dash in the 2.5.109 entry is gone). Facts are unchanged - same releases, same changes; this is a formatting/copy-tightening pass, not a rewrite of history. - Guide truncation fixed too.
GUIDE_2567 + HOWTO_TEXT + GUIDE_TEXT totalled ~37.5K chars against the dialog's fixed 32768-byte combine buffer, so snprintf silently cut off the oldest ~4.7K (part of the Feature Guide never displayed). The buffer is now 65536 and the cleaned guide (~34.5K combined) fits whole, with headroom for future entries. Regression coverage: new server/tests/client/guide-format-test.sh decodes the actual guide strings from src/tray.c and fails if any line ending is a bare LF, any heading is not at a line start or lacks its preceding blank line, any heading is not followed by a - change, any line carries a double space, any character is non-ASCII, the first heading is not the current APP_VERSION, or the combined texts outgrow the dialog buffer (mutation-verified: reintroducing the blob makes it fail). Version metadata synced to 2.5.118 in src/cloud.h (APP_VERSION), all three .rc files, the update manifest and all four EXEs. Client-only release; the server stays 2.5.138. Nothing for Austin to configure.
- Threat Guide search + shareable links. The
/threat-definitions guide now opens with a search box that live-filters all 84 threat cards as you type (group sections with no matches hide; a match count and a no-results message with a clear-search link are shown). The page is unchanged with JavaScript off. Every threat card already carried a stable anchor id — opening the guide with a #anchor now also scrolls the card into view past the sticky header and highlights it, so links like /threat-definitions#ransomware land exactly on the threat. - "Is this link safe?" checker (
/link-checker). A free public tool: paste a link, get a Safe / Caution / Dangerous / Unknown verdict with the specific reasons. The submitted URL is analyzed as text only — the server never fetches or resolves it, so there is no SSRF surface. Evidence comes from the server's own threat data (the merged intel-feed blocklists in threats.json, feed_indicators, fleet IOC broadcasts, and client-reported iocs) plus structural signals mirroring the Windows client's URL engine (src/urlml.c): raw-IP hosts, punycode, suspicious TLDs, URL shorteners, excessive subdomains, credential/pressure keywords, the "@" userinfo trick, hyphen-stuffed domains, non-standard ports, and http sign-in pages — plus brand impersonation (a famous brand token as a host segment on a domain that isn't the brand's). "Safe" is only claimed for a well-known domain with zero hard signals; a clean-but-unknown domain honestly returns "Unknown". The page carries a plain "guidance, not a guarantee" disclaimer, and the endpoint is rate-limited per IP (30/hour). - Live protection stats band. New public endpoint
GET /api/protection-stats (5-minute in-memory cache) returns real counts only — registered devices (installs), the monotonic threats-stopped counter, detections recorded this calendar month, and the merged blocklist size (feed domains + IPs). The homepage shows them in a stats band that stays hidden unless real numbers arrive — no placeholder stats are ever rendered. - Comparison table on
/subscribe. AEAntivirus vs Microsoft Defender vs "Typical free antivirus" across 11 features, marked honestly: Defender is credited with real-time protection and behavioral/ML detection; AE's edges are ransomware rollback, DNS-level blocking, built-in parental controls, a password manager, one license across Windows/Linux/Android, human technician phone support, the 200+ add-on catalog, the ransomware warranty, and no nag popups. A footnote dates the comparison and notes Defender's nuances (Controlled Folder Access can't restore encrypted files; web protection is browser-scoped; Family Safety is a separate app). - Smart download page.
/download installer cards now carry data-plat; a small script reads the visitor's user agent, moves that platform's card first, outlines it, and adds a "✓ For your device" badge. Presentation only — all cards remain visible/downloadable, and the 2.5.110 expired-license gating is untouched. Help/FAQ gained a link-checker entry and a note about Threat Guide search + shareable anchors; the footer Support column and the sitemap link to /link-checker. Server-only release; the Windows client stays 2.5.114.
Version 2.5.117
- Quarantine restore, repaired (Windows). The owed fix for Austin's report ("restore failed - check that the original location is writable", and Delete appearing to do nothing). Root causes found in the pre-2.5.117 code: (1) restore did a bare
MoveFileA from the vault - which lives under %PROGRAMDATA%/%LOCALAPPDATA% on C: - so restoring a file whose original location was on another drive always failed with ERROR_NOT_SAME_DEVICE (17), with no copy fallback (the quarantine-IN path had one; restore never did); (2) restore deleted any existing file at the destination first and only then tried the move, so a failed move (cross-volume, or the destination was locked/running so the delete itself failed and the move died with ALREADY_EXISTS) both failed the restore AND could destroy the file that was already there; (3) the record lookup read only the current mode's quarantine.db while the .quar file lookup already spanned both the tray (%LOCALAPPDATA%) and service (%PROGRAMDATA%) dirs, so a service-quarantined item could be un-restorable from the tray with "record not found"; and (4) when a restore failed through a path that left the detailed error empty, the tray fell back to the generic "Check that the original location is writable." text Austin saw. Restore is now a portable, unit-tested core (AV25117-TESTABLE-QUAR in src/aeantivirus.c): recreate the parent folder tree, move an existing destination aside to <dest>.aeabak (never delete it first), try a same-volume rename, fall back to a chunked copy with a size check (the cross-volume route), delete the .quar copy only after the restored bytes are verified in place, and roll the original destination back untouched on any failure. Every failure path in restore and delete now sets a specific quarantine_last_error() reason, and record find/remove span both modes' DBs. The tray also no longer calls engine_quarantine_delete() after a successful restore (restore already removes file + record; the extra call cleared the error string as a side effect), and the FP verdict paths only delete when restore actually failed, preserving the 2.5.90 "FP always clears the record" rule. - Quarantine Delete - what was actually wrong. In current (2.5.116) source the Delete wiring is correct - button
IDC_QUAR_DELETE -> confirm -> engine_quarantine_delete -> file + record removed, and ui_msgbox has been the standard MessageBoxA (correct IDYES contract) since 2.5.80 - and 2.5.117 pins that wiring with source guards. The dead-Delete symptom traces to Austin's long-stuck older build (<=2.5.113 era and earlier): pre-2.5.80 builds used a custom themed message-box whose per-module static state broke its buttons, so the Yes/No confirmation could fail to return IDYES and Delete silently did nothing; a delete could also strand on the single-mode record/file mismatch above. Both are gone in current code, and delete now goes through the same tested portable core as restore (missing file -> the stale record is dropped, as before). - Fleet scheduled scans, client half (Item 8b). Server 2.5.137 already delivers
policy_scan_schedule ({enabled, type:'quick'|'full', days:[0-6, 0=Sunday], time:'HH:MM'}) in the heartbeat policy; the client now consumes it. cloud.c parses and re-sanitizes the object (AV25117-TESTABLE-SCHEDPOL, same fallbacks as the server) and persists it via cfg_set as ssched_managed / ssched_enabled / ssched_type / ssched_days / ssched_time. Enforcement lives inside the existing schedscan.c scheduler - no parallel scheduler: a managed + enabled policy schedule takes precedence as the fleet schedule in schedscan_check(), the scheduler thread also runs when only the policy schedule is active (local toggle off), and an absent or disabled policy leaves the legacy local daily/weekly schedule exactly as before. The slot id is the local date, persisted as ssched_last_slot after a scan actually runs, so a slot fires at most once. Catch-up semantics, stated plainly: none beyond a 15-minute firing window after the scheduled time (it absorbs the 60-second poll and brief sleep/wake delays) - a slot missed while the PC was off is skipped, not fired late. Gates: a scan already running defers within the window (engine claim, slot unmarked), and a license expired past grace never starts a scheduled scan (cloud_protection_allowed()). Type maps to the existing remote-command entry points: quick -> engine_scan_quick(), full -> engine_scan_full(). - Version metadata synced to 2.5.117 in
src/cloud.h (APP_VERSION), all three .rc files, the update manifest and all four EXEs. The server stays 2.5.137 (only server/views/home.js What's-New rotation and this README change server-side). Client coverage: new server/tests/client/quar-test.c + .sh (27 checks: byte-identical round-trips via rename and forced-copy paths, deleted parent tree recreated, delete removes file + record, failed restore onto an existing destination loses nothing and retries clean, every failure carries an error string), polsched-test.c + .sh (35 checks: schedule parsing/sanitizing, due-at-slot / not-before / window / no-late-catch-up / never-twice / next-chosen-day / disabled / days-subset, and a config write/reload round-trip of the schedule + last slot), and quar-sched-guard-test.sh (24 source guards: Delete wiring to the engine, restore/delete through the portable core, old destructive restore gone, dual-DB lookup/removal, heartbeat -> config -> schedscan wiring, quick/full mapping, license gate).
- Changelog page fixed — every release is listed again. The /changelog page parses this README's "New in" sections, but its parser only recognized the plain
## New in X.Y.Z form. Since 2.5.110 the entries are headed ## New in Server X.Y.Z / ## New in Client X.Y.Z, so the page silently froze at 2.5.109 — releases 2.5.110 through 2.5.116 never appeared. The parser now accepts the optional Server/Client prefix (the boot-time README staleness check had the same flaw and is fixed too), and sections that share a version number — common now that the server and client streams both use 2.5.x — merge into a single card instead of rendering twice or as an empty duplicate. Server-only release; the Windows client stays 2.5.114.
Version 2.5.116
- Protection pause, rebuilt to spec (Windows). The tray's Pause protection submenu now offers exactly 15 minutes, 1 hour, and Until restart, with Resume protection in the tray menu while paused and a new Resume protection button on the Status page. A pause surface already existed (15/30/60 minutes + indefinite, a
paused config flag that gated the real-time watcher), but it was largely cosmetic for the other modules and its deadline was a GetTickCount value that could not survive a restart honestly. 2.5.116 fixes the gaps instead of adding a parallel system: pausing now snapshots the user's own per-module states (pause_prior_fw / pause_prior_wp / pause_prior_ar in config.ini) and then turns the Firewall, Web Protection, and Advanced Real-Time Protection off through their own setters (fw_set_enabled(0) clears the firewall rules, wp_set_enabled(0) clears the hosts section) - the same code paths the UI toggles and fleet policy use. Real-time scanning stays gated by engine_paused() as before. - Resume restores your choices exactly. Every resume path (tray menu, Status button, the Protection-page real-time toggle, timed auto-resume, cold-start reconciliation) goes through one restore: each module is set back to its captured prior state via the same setters, so a module the user had manually OFF before pausing comes back OFF, not ON (Austin's standing rule: manual choices stick). Re-pausing to extend a pause never re-snapshots the already-disabled modules as the "prior" state.
- Wall-clock deadline + cold-start rules. The deadline is now a persisted epoch (
pause_until_epoch in %PROGRAMDATA%\AEAntivirus\config.ini); the legacy pause_resume_tick is still honored for a pause in flight across the upgrade and cleared by every resume. A deadline that expired while the app was closed resumes on the next start (never still-paused, never extended); auto-resume also fires live at the deadline via the existing tray refresh/heartbeat timers and the engine's own engine_paused() checks, restoring protection with no user interaction. Until restart = deadline 0 plus the session token of the process that set it (pause_session): it survives UI refreshes and window close/reopen, and is cleared by engine_pause_on_startup() on the next cold start of the tray/service after a full exit (or a reboot). - Status honesty + policy lock. While paused the Status page shows "Protection paused - Attention needed" (amber status circle, never green and never "all protection running normally"), the tray icon gains an amber pause badge and a "protection paused" tooltip, and heartbeats keep reporting the truth through the existing fields -
protection_paused / paused_until (now taken directly from the epoch deadline) and the module states, which read OFF for the paused modules because they really are off (no protocol change). If fleet policy locks protection toggles (policy_lock / lock_settings), Pause is refused: the tray menu and the Status/Protection toggles show the standard message box "Protection settings are managed by your administrator." and the engine refuses the pause as a backstop. Server stays 2.5.135 - nothing server-side changed. - Version metadata synced to 2.5.116 in
src/cloud.h (APP_VERSION), all three .rc files (src/version.rc + the generated x86/x64 copies), the update manifest and all four EXEs (x86 + x64, installer + updates). Client coverage: new server/tests/client/pause-test.c + pause-test.sh (31 checks: deadline per duration, expiry incl. expired-while-closed, cold-start actions for timed and Until-restart pauses, snapshot guard, and a config write/reload round-trip proving paused state, deadline, session and per-module priors survive and restore exactly) and pause-guard-test.sh (23 source guards: exact menu durations, setter-based disable/restore, epoch key + cold-start call, policy-lock refusal text, status phrase + Resume control + tray-icon wiring, heartbeat fields).
- Custom plan: automatic volume discounts. Austin's tiers, volume style — once the device count crosses a threshold, ALL devices bill at the discounted rate: 11–24 devices → 10% off ($4.49/device), 25–49 → 20% off ($3.99/device), 50+ → 30% off ($3.49/device) off the $4.99/device base. The discount applies automatically everywhere the Custom quantity is set or changed: at checkout (the count chosen during subscribe), on every self-service device-count change in My Account, on registration auto-grow, and on the removal/admin shrink paths — crossing up applies or raises it, dropping back down reduces or removes it, so a customer never keeps a discount their count no longer qualifies for. Existing Custom subscribers pick up the right tier at their next quantity sync (the adjust path reconciles even when the count itself doesn't move). Proration stays Stripe-default; other plans' subscriptions are never touched.
- Mechanism: server-managed Stripe coupons — zero dashboard setup. Three coupons with fixed ids (
aeav-volume-10/20/30, percent-off, duration forever) are created via the Stripe API by the server the first time each is needed (ensureVolumeCoupon(), retrieve-then-create, memoized) and reconciled onto the subscription by reconcileCustomVolumeDiscount(), which runs inside the stripeSetCustomQuantity() funnel every quantity change already passes through, plus the checkout-session creation and the checkout webhook as backstops. Reconciliation is idempotent — when the subscription already carries the right coupon it makes no Stripe write at all — and it only ever manages coupons it owns: a foreign coupon (e.g. a promo code, which wins the checkout session when the customer types one) is left in place. A discount failure never blocks a quantity change; it is logged and audited for the next sync to retry. - Customers see the discount before and after buying. The Custom plan card (shared
priceCards() component) lists the ladder in its features; the subscribe page device picker shows the effective per-device rate and discount for the selected count as it changes ("$3.99/device — 20% volume discount · $119.70/mo"); the My Account custom-plan summary shows the effective rate, a "N% volume discount applied" badge, and the discounted monthly total, and the Add-a-device button is priced at the effective rate. The 2.5.115 price-fit work is untouched (amount/suffix split unchanged). Regression coverage in server/tests/server/v25116-volume-discount-test.js (stubbed Stripe, real requests: checkout at 5/15/30/60, promo precedence, webhook creation, count changes up/down across every threshold, idempotent same-tier syncs, existing-subscriber reconcile, registration auto-grow crossing, card/picker/account display). Server-only release; the Windows client stays 2.5.114.
Version 2.5.115
- Announcement popups (Windows). The client now honors the server's new
announce command: when your admin (AEAntivirus / your IT admin) sends an announcement to your PC, it appears once as a standard Windows message box (announcement title + message, OK button, information icon — no custom dialog). Four-place rule (client half): announce added to the cmd_parse_one whitelist in src/commands.c (with full payload validation: strict id + canonical base64 title/body, decoded in the pure testable block), dispatched in commands_dispatch, and routed in engine_run_command (src/aeantivirus.c) to the same handler — the display lives on the tray/UI side, following the remote-lock pattern: in the user (tray) process the box is shown directly; in the service process the announcement is persisted to announce_pending.txt under the program-data folder and the tray shows it on its next notify poll (and at startup), so a service-side pickup can never swallow it. - Shown at most once per PC. Seen announcement ids persist in
config.ini (announce_seen, bounded to the newest 50 ids) under %PROGRAMDATA%\AEAntivirus. The id is marked seen BEFORE the message box opens (the 2.5.114 modal discipline — a modal loop keeps dispatching timers, and an announcement marked only after dismissal could stack or repeat); an already-seen id is skipped silently and reported as handled. The pending file is likewise deleted before the box opens. - Version metadata synced to 2.5.115 in
src/cloud.h (APP_VERSION), all three .rc files (src/version.rc + the generated x86/x64 copies), the update manifest and all four EXEs (x86 + x64, installer + updates). Client coverage: server/tests/client/cmd-test.c gains announce parse cases (valid payload decodes intact; empty/zero/leading-zero/non-numeric ids, bad base64, missing/extra fields, and over-long payloads rejected; unknown and over-long types still rejected) plus seen-set helper cases (contains/add, dedupe, 50-id bound with oldest dropped), and server/tests/client/announce-test.sh asserts at source level that the mark-seen call precedes the MessageBoxW call in the display path and that the pending file is removed before showing.
- Pricing cards: the Custom plan's price fits its card. The subscribe page rendered the Custom plan's price as one string at the 38px display size —
$4.99/device big, with only the trailing /mo shrunk — so it ran past the card's right edge and the rest was cut off (Austin's screenshot). The shared pricing-card component (priceParts() + priceCards() in server/server.js) now splits every plan's price into the amount, which keeps the big display size in its own <span class=amt>, and the suffix (/mo, /device/mo), rendered smaller in <span class=per> — the standard pricing-card pattern, applied uniformly to all plans so any future long suffix fits too. The .price-tag CSS is now a wrapping flex row (baseline-aligned, wraps only between amount and suffix, both spans nowrap) with a clamp(28px,3.2vw,38px) display size and max-width:100%, so a price can never overflow its card on desktop or mobile. Card design otherwise unchanged. Server-only release; the Windows client stays 2.5.114.
Version 2.5.114
- My Account: Delete license, Remove device, and Transfer — customers can manage their own licenses and devices. Every license card on /account now has a Delete license button, every device listed under a license has a Remove button, and each license with at least one device gets a Transfer to a new device button (with a device picker when the license has several). All three actions are ownership-verified server-side — another account's key or install returns 403 and changes nothing.
- Delete license = soft delete + immediate subscription cancellation. The license row is kept as status
deleted with a deleted_at stamp (admin and history stay intact), it disappears from My Account, licenseStateForClient() reports it as a bad/revoked state, registration with the key is refused, and every install under it is deactivated and tombstoned. If the license has a Stripe subscription it is canceled immediately (stripe.subscriptions.cancel — billing stops today, no refund), never the account's separate add-on subscriptions. - Remove device = deactivate + tombstone + immediate subscription cancellation. The install row is deleted (seat freed) and a
revoked_installs tombstone keeps that install_id and device fingerprint refused on heartbeat (403 install_revoked) and on re-registration under that license; a brand-new install can still take the freed seat. The license's Stripe subscription is canceled immediately, the same as Delete. - Transfer = the device-swap path, billing untouched. For "I got rid of a device but have a replacement of the same type": the old install is deactivated and tombstoned exactly like Remove, the seat frees up, and the customer installs on the new device with the same key — the license stays active and no Stripe call of any kind is made; the subscription and billing continue unchanged. Same-platform is enforced by the existing registration scope check ("This Windows license key cannot be used on Linux").
- Stripe failure handling: if a subscription was already canceled in Stripe, that counts as success. If the immediate cancel call fails for another reason, the local delete/remove still completes, the failure is logged + audited + admin-alerted, and the API response carries
stripe_cancel: "failed" ("canceled" / "already" / "none" otherwise — "none" for trial and admin-granted licenses with nothing to cancel). The customer.subscription.deleted webhook now treats an already-deleted license as an idempotent no-op instead of flipping it back to canceled. - Endpoints:
POST /api/account/licenses/delete, POST /api/account/licenses/transfer, and the upgraded POST /api/account/installs/remove (403 for another account's install; response gains stripe_cancel). Help & FAQ gains "How do I remove a device or delete a license?". Regression coverage in server/tests/server/v25114-license-delete-test.js (seeded server + stubbed Stripe: cancel called for the right subscription id on delete and on remove, zero Stripe calls on transfer and on trial delete, already-canceled and cancel-failure paths, webhook idempotency, cross-platform register refusal, ownership 403s). Server-only release; the Windows client stays 2.5.114 (the shared number is a coincidence — server and client versions are independent).
- Expired-license payment prompt no longer stacks: the daily "Subscription Expired" reminder (2.5.104) is modal, but a modal dialog's message loop still dispatches WM_TIMER to the main window — and IDT_REFRESH fires every 5 seconds. The 24-hour throttle stamp (
cloud_license_prompt_mark()) was only written after the user answered, so cloud_license_prompt_due() stayed true the whole time the prompt sat open and every timer tick spawned another identical dialog (Austin's screenshot showed five stacked copies). The reminder now stamps the throttle the moment it appears — persisted even if the prompt is never answered or the app exits with it open — and a single-instance guard (g_licPromptOpen, the same InterlockedExchange single-flight pattern as g_syncBusy) makes a second concurrent copy impossible. Audit of the other prompts: the threat-verdict dialog already has the 2.5.107 per-path open guard; ui_ask_yesno() callers each block their own single worker thread until answered, so none can re-fire its own prompt while it is open; the netmon / remote-lock / installer timers open no dialogs. The license prompt was the only instance of the pattern. Regression coverage in server/tests/client/licprompt-stack-test.sh. Client-only release; the server stays 2.5.111.
Version 2.5.113
- Homepage "New in" section trimmed: the homepage now shows only the newest 3 update blocks; all older update notes (back to 1.58.4) live in a single "See older updates" collapsible on the same page, so the homepage stays short without losing the history. This also replaces the stack of nested "Show older updates" collapsibles that past releases had accumulated. No update wording changed, and the full history is still on /changelog. Standing rule from here on: keep the newest 3 updates visible and move the rest into the collapsible (noted in a comment at the top of the section in server/views/home.js). Server-only release; the Windows client stays 2.5.114.
- Server-anchored grace clock: the client now parses
grace_until from heartbeat and registration responses and combines it with any locally recorded anchor through a single rule (grace_anchor_apply(), pure C, unit-tested): the effective grace end can only move EARLIER, never later. The old local first-sighting now + 72h remains only as a legacy fallback when no anchor exists anywhere. Two latent clock-reset bugs fixed along the way: a successful heartbeat used to zero the grace unconditionally (re-arming the local fallback on every cycle, so the clock slid forever), and (re)registration did the same — both now clear grace only when the license is actually valid again (active / trial / standalone). Protection, the payment prompt, update blocking, and the banner all read the same anchored cloud_grace_until(), so the banner now shows the true plan-anchored end date. - The grace deadline is in the registry identity backup:
cloud_backup_identity() also writes LicenseGraceUntil (including 0, so a reactivation-cleared grace can't be resurrected from a stale value), and reg_restore_identity() restores it on startup when the config lost it — applying the same earlier-wins rule, so even an offline reinstall that wipes %PROGRAMDATA%\AEAntivirus\config.ini can't reset the clock. - Regression coverage in
server/tests/client/grace-anchor-test.c (extracts the AV25113-TESTABLE-GRACE block from src/cloud.c; covers adopt / can't-extend / can-shorten / no-candidate / reinstall-restore scenarios).
# AEAntivirus Client 2.5.112
Version 2.5.112
- License keys now live only on My Account: the download page no longer shows a "Your license key" card. In its place is one slim pointer — "Installing? Your license key is on your My Account page — you'll enter it when the installer asks." Trial-key provisioning is unchanged: the key is still created server-side when a signed-in user visits /download; it simply displays only on /account now. Anonymous visitors see the download page exactly as before.
- Every key on My Account is labeled with its platform: each license on /account now shows a "Platform: ..." line — the key's own platform when it has one (Windows / Android / Linux / any device), otherwise the platform(s) of the devices actually registered under that key, otherwise "Not used on a device yet".
- Add-on card icons fixed on /account: the three newest add-ons (Wi-Fi Security Scanner, Scam Call Blocker, Child-Safe Search) — and 51 more — had no entry in the account page's icon map, and the map's fallback was a raw escape string, so their cards showed the literal text of an escape sequence instead of an icon. All 127 catalog add-ons now have a real icon, and the fallback is a real emoji character, so a future add-on can never render as escape text again.
- Admin Add-ons sections can no longer fail silently: the Command Center table, the paid add-on catalog, the service bookings and the phishing campaigns loaded through shared, fully-swallowed error paths — one failed fetch (or an HTML error page where JSON was expected) emptied several sections at once, left "Phishing training campaigns" stuck on "Loading...", and the lazy loader stayed marked as loaded so revisiting the tab never retried. Each section now loads independently, checks the HTTP status before parsing, and writes a visible error with a Retry button into its own table on failure; the loader registry clears a loader's loaded mark when it throws or rejects so the next visit retries; and the phishing-campaigns and service-bookings endpoints now answer JSON errors instead of HTML 500s. Verified against a seeded database (active license + install, add-on config, price override, paid add-on, campaign, booking): all five endpoints return 200 with full data (command-center 227 add-ons = 127 paid + 100 free; catalog/prices 127). Server-only release; the Windows client stays 2.5.114.
- Hosts guard no longer flags AEAntivirus's own cleanup: since 2.5.109, the hosts-file monitor's self-protection automatically strips hosts entries that block aeantivirus.com — but it rewrote the file without telling the hosts guard (guard.c) the write was ours. The guard's 30-second poll then saw its baseline change and popped "The hosts file was modified by another program (not AEAntivirus). Restore it to AEAntivirus's last-known content?" — and clicking Yes restored the pre-cleanup file, putting the aeantivirus.com blocks right back. The self-protection write in
hostsmon_selfprotect() now uses the same guard handshake as webprotect (hosts_guard_self_write_begin() before the write, hosts_guard_note_mine() after it — including on the write-failure path, which also clears the self-write flag and re-baselines the guard to the cleaned file, so the ordering is safe even when the cleanup runs before the guard's first-ever baseline check: note_mine adopts the cleaned file as the baseline). The same missing handshake was found and fixed in the tray's Focus Mode writer (r2b9_focus_apply()), which rewrites the hosts file's Focus section the same way. All other hosts-path users were audited: webprotect already handshakes; advprot/integrity only read the file; guard's own restore runs inside the guard thread. Regression coverage in server/tests/client/hosts-guard-handshake-test.sh. Client-only release; the server stays 2.5.109.
# AEAntivirus Client 2.5.111
Version 2.5.111
- Threat Guide expanded a fourth time — 84 threat types (was 70): the
/threat-definitions page (server/views/definitions.js) adds 14 entries in the established four-part format (what it is / how it gets in / how AEAntivirus stops it), each tied to a real engine or stated honestly as awareness: ClickFix / fake CAPTCHA scams (fileless detection is the backstop), remote-access scams, malicious game cheats & mods, cracked software & keygens, OAuth consent phishing, inbox-rule attacks, SEO poisoning, browser notification scams (with the exact steps to revoke notification permission), gift card scams, government impersonation scams, job & task scams, fake charity scams, Bluetooth attacks (honest: the fix is patching), and cloud account hijacking (2FA + breach monitoring). - Two new guide sections: "Scam red flags" — the ten universal moves (urgency, secrecy, gift-card payment, too-good-to-be-true, unexpected contact, off-platform pressure, install-my-screen demands, move-money-to-protect-it, and more) with the one counter-move that defeats them all; and "If you already clicked or paid" — an ordered damage-control list: call the bank/card issuer now, gift-card company with card + receipt in hand, passwords changed from a clean device (email first, then 2FA and session revocation), Full Scan if anything was installed or remote access was given, credit freeze for shared personal details, and reporting to reportfraud.ftc.gov and ic3.gov.
- Quiz grows from 6 to 10 scenarios: "Test yourself: scam or legit?" adds fake-CAPTCHA (ClickFix), boss gift-card demand, remote-access request, and one legitimate bank-letter scenario; scoring thresholds updated (perfect 10, good 7+). The anchor index is generated from the groups array and now covers all 84 entries. Server-only release; the Windows client stays 2.5.113.
- DNS filter can no longer strand the PC offline: fixed a self-poisoning bug in the DNS filter. If a crash or restart left a network adapter pointing at the filter's own 127.0.0.1 redirect and the one-shot startup recovery missed it (adapter not up yet), activation saved "127.0.0.1" as the adapter's *original* DNS — the proxy then forwarded every query to itself, answered SERVFAIL for everything, and the health probe counted that SERVFAIL as healthy, so no website would load (DNS_PROBE_FINISHED_BAD_CONFIG) while the app ran normally. Now: loopback/unspecified addresses are never saved, restored, or used as upstreams (
dnsf_sanitize_dns_list); configs already poisoned by older builds are migrated to DHCP on read; the health probe validates the response (matching ID, QR set, RCODE != SERVFAIL — NXDOMAIN still counts as healthy) so a dead upstream chain trips the watchdog's fail-open restore; and stale-DNS recovery re-runs ~30s and ~120s after startup (only while the filter is inactive) to catch adapters that come up late. Regression coverage in server/tests/client/dnsf-guard-test.c (29 checks). Client-only release; the server stays 2.5.109.
# AEAntivirus 2.5.109
Version 2.5.110
- Expired licenses can no longer be circumvented by reinstalling — download gating: a logged-in account whose licenses are ALL in a bad state (expired / past_due / canceled) no longer gets installer downloads.
GET /download renders a reactivation page ("Your license has expired — update your payment…", linking to /account) instead of the installer cards, and both installer file routes (/download/file/:name and /download/AEAntivirus-Setup.exe) redirect those accounts to /download instead of serving bytes or proxying to the update channel. Accounts with any active/trial license — and anonymous visitors (prospects with no license to expire) — are unaffected, and no trial key is provisioned for a gated account. Helper: accountLicensesAllBad() in server.js; new downloadExpiredView() in server/views/download.js. - Heartbeat (and registration) responses carry the authoritative grace anchor:
licenseStateForClient() now adds grace_until (epoch seconds = current_period_end + 72h) whenever the license is expired / past_due / canceled — or a trial past its end — and a period end exists. The 72h grace is thereby anchored to the plan's end date on the server, not to whenever a client happens to notice. Omitted for good-state licenses and when current_period_end is NULL. - Help & FAQ: new "What happens when my license expires?" entry (Account and billing) documenting the anchored 72h grace, the download pause, and instant reactivation on payment. (Terms of Service does not describe grace/reinstall behavior — checked, no ToS change.)
- Regression coverage in
server/tests/server/v25110-grace-download-test.js (25 checks: anchored heartbeat values for expired/past_due, omission for active and NULL-end, gated page + both file routes, active + anonymous flows). Inline-script regression passes.
- No more firewall prompts for Windows itself: known Windows components under the Windows folder (explorer.exe, svchost.exe, dwm.exe, RuntimeBroker, SearchHost, and two dozen more) are now allowed automatically instead of prompting. A same-named program living outside the Windows folder still gets asked about.
- Firewall decisions actually stick now: remembered Allow/Block choices moved from the config file — whose ~1KB value cap silently dropped every decision after roughly 35 programs, so the same apps kept asking forever — to a dedicated
appfw_rules.txt file (one decision per line, up to 256 programs). Existing decisions are migrated automatically on first run, and a failed save is now logged instead of silent.
# AEAntivirus Client 2.5.109
Version 2.5.109
- Threat Definitions renamed to Threat Guide: the footer link, page title, and H1 at
/threat-definitions now read "Threat Guide" — same page, same URL, all 70 threat definitions, the "What our alerts mean" decoder, and the scam-or-legit quiz unchanged. The Help & FAQ pointer and the homepage "New in" note use the new name too. Server-only release; the Windows client stays 2.5.110.
# AEAntivirus 2.5.108
- Self-protection for aeantivirus.com: the hosts-file monitor now strips any hosts entries that block aeantivirus.com (or its subdomains) to a non-routable IP (0.0.0.0, 127.x, 10.x, 172.16-31.x, 192.168.x). Checked on every hosts change and at startup — only our domain is touched, all other entries are left alone. Keeps updates, license checks, and heartbeats working even if an aggressive blocklist nukes our domain.
- Parental controls no longer block aeantivirus.com: fixed a bug where the DNS parental-controls check overrode the infrastructure exemption for our domain. aeantivirus.com and its subdomains are now always exempt from both blocklist and parental filtering.
# AEAntivirus Client 2.5.108
Version 2.5.108
- Threat Definitions page expanded a third time:
/threat-definitions now defines 70 threat types (was 56). Fourteen new entries in the same plain-language format (what it is, how it gets in, how AEAntivirus stops it — with honest "awareness is the defense" copy where that is the truth): smishing, vishing, deepfake & AI voice-cloning scams, fake browser update, LNK shortcut malware, malicious disk images (ISO/IMG), DLL sideloading, logic bomb, browser-in-the-browser phishing, MFA fatigue (push bombing), SIM swapping, juice jacking, package delivery scam, and marketplace scams. Two new sections: "What our alerts mean" — 14 real client detection names (AI.Heuristic.Suspicious, ML.Behavior.Ransomware, Ransomware.Behavior.Canary/Sabotage, Fileless.Threat, Behavior.LOLBin, Behavior.OfficeMacro, Defense.HostsFile, Firewall.AppBlocked, Keylogger.Detected, Clipboard.Hijack, Rootkit.Hidden, and more) translated into what triggered them and what to do — and "Test yourself: scam or legit?", an interactive 6-scenario quiz with instant feedback and a final score. The anchor index at the top of the page covers all 70 entries. Server-only release; the Windows client stays 2.5.110.
# AEAntivirus 2.5.107
- License key survives updates: the license key, install ID/token, server URL, and account email are now mirrored to
HKLM\SOFTWARE\AEAntivirus in the Windows registry on every save. If the config file is ever lost or wiped during an update, the client automatically restores them from the registry at startup — no more re-entering your license key.
# AEAntivirus 2.5.103
Version 2.5.107
- Threat Definitions page expanded again:
/threat-definitions now defines 56 threat types (was 42). Fourteen new entries in the same plain-language format (what it is, how it gets in, how AEAntivirus stops it, tied to real protections): watering hole attack, typosquatting, session hijacking / cookie theft, credential stuffing & password spraying, firmware / UEFI malware, router & IoT malware, evil twin / rogue Wi-Fi, crypto wallet stealer, point-of-sale (POS) malware, clickjacking, social engineering, sextortion scam, investment scam / "pig butchering", and malicious browser extension. Three new sections: "How an infection unfolds" (the five-stage attack chain and where AEAntivirus breaks it), "Myths vs facts" (seven security myths corrected), and "Security terms, in plain English" (a 15-term glossary). The anchor index at the top of the page covers all 56 entries.
# AEAntivirus 2.5.106
- No more stacked threat popups: the one-click verdict dialog now opens at most once per file — if a dialog is already open for that file, a new one won't queue behind it. Clicking "Decide Later" snoozes popups for that file for 1 hour instead of doing nothing (previously the same file could re-trigger a popup immediately, stacking dozens of dialogs). Once you give a verdict (False Positive or Confirmed Threat), you won't be asked again for that file this session.
- Expired licenses blocked from updates: update checks now return immediately for expired/past_due/canceled licenses — even during the 72-hour grace period (grace keeps protection running, but new builds require a current subscription).
- Grace end date on the banner: the expired-license banner now shows when the 72-hour grace period ends (e.g. "grace period ends Oct 14 at 9:26 PM").
# AEAntivirus Client 2.5.106
Version 2.5.106
- Threat Definitions page expanded:
/threat-definitions now defines 42 threat types (was 26). New entries, in the same plain-language format (what it is, how it gets in, how AEAntivirus stops it, tied to real protections): exploit / zero-day attack, living-off-the-land (LOLBins), process injection, credential dumping, screen locker, web skimmer / formjacking, malvertising, drive-by download, spear phishing & business email compromise, QR-code phishing (quishing), tech support scam, DNS tunneling / DGA, polymorphic malware, USB / autorun malware, supply-chain attack, and double-extortion ransomware. Two new sections: "Signs your PC may be infected" (symptom checklist pointing to a Smart Scan) and "What to do if you think you're infected" (numbered response steps, including Ransomware Rollback). The anchor index at the top of the page covers all new entries.
# AEAntivirus 2.5.105
- Batched hosts-file alerts: when many hosts mappings change at once (e.g. a blocklist app adding 20 entries), you now get one summary notification listing up to 10 changes instead of a toast per mapping. Throttled to at most one hosts alert every 5 minutes. Per-change threat logging is unchanged.
# AEAntivirus Client 2.5.105
Version 2.5.105
- New Threat Definitions page: a "Threat Definitions" link in the site footer (Support column) opens
/threat-definitions, defining every threat type AEAntivirus detects — virus, worm, trojan, ransomware, spyware, keylogger, rootkit, backdoor/RAT, adware, PUP, phishing, cryptominer, scareware, stalkerware, infostealer, credential stealer, clipboard hijacker, dropper/loader, wiper, botnet malware, fileless malware, macro malware, browser hijacker, DNS hijacker, bootkit/MBR malware, and banking trojan — in plain language: what each one does, how it usually gets in, and how AEAntivirus stops it, tied to real protections (behavioral ML, ransomware rollback, Boot Guard, hosts/DNS monitoring, phishing feeds). The Help & FAQ page now points to it, and it is in the sitemap.
# AEAntivirus 2.5.104
- Expired-license banner: as soon as your subscription expires, a yellow banner appears at the top of the main window with an "Update Payment" button — no waiting for the grace period to end. During the 72-hour grace it reads as a warning; after grace it confirms protection is off.
# AEAntivirus Client 2.5.104
Version 2.5.104
- Expired main license disables all paid add-ons: the heartbeat now returns an empty add-on list when the main license is expired/past_due/canceled. Clients gate paid features (VPN, DDoS protection, identity monitoring, etc.) on that list, so expired users lose access to every paid add-on immediately. (Per-add-on cancellation already worked —
paid_addons flips to 'canceled' via the Stripe webhook and the list filters to status='active' — only the main-license-expired case was missing.)
# AEAntivirus Client 2.5.110
- Expired subscription payment reminder: when your subscription expires and the 72-hour grace period ends, AEAntivirus now shows a daily reminder with an "Update Payment" button that opens your account page directly — no more silent protection shutoff.
- Server policy respects your protection toggles: fleet policy can still turn protection OFF for security, but it will no longer re-enable protection you manually disabled unless your admin has locked the toggles.
# AEAntivirus 2.5.102
Version 2.5.103 (admin-only)
- Successful payments reactivate licenses immediately: the
invoice.payment_succeeded webhook now flips a past_due/expired/canceled license back to active right away (and refreshes current_period_end from the invoice) instead of waiting for a customer.subscription.updated event that may never fire if Stripe doesn't see a status change.
# AEAntivirus Client 2.5.107
Version 2.5.102 (admin-only)
- ML training status no longer hangs: the trainable-verdict count is now served from a 5-minute cache instead of scanning the verdict log on every request, so the admin ML panel can't get stuck on "Model training: loading..." when the table is large or locked.
# AEAntivirus 2.5.101
Version 2.5.101 (admin-only)
- Detection history preserved on verdict: marking a detection as FP/TP now marks it
resolved_at instead of deleting the row, so the "Detection trends — last 30 days" graph keeps its history (previously the graph went flat as verdicts wiped the underlying rows). The Detections list shows only unresolved items; resolved ones stay in the database for trends and CSV history. Applies to both admin and client (one-click) verdict paths.
# AEAntivirus 2.5.99 / Client 2.5.100
Version 2.5.100 (admin-only)
- ML training verdict count fix: the admin ML panel now shows trainable verdicts alongside the raw counter (e.g. "137 new verdicts (42 trainable)"), so the retrain recommendation matches what Train can actually use. Client verdicts now log to the training history like admin verdicts do, and verdicts without recognized signal tags no longer inflate the retrain counter.
Version 2.5.98 (admin-only)
- Ransomware Rollback events cleanup: the admin Ransomware Rollback events endpoint no longer returns
watcher_started noise events, so the Ransomware Rollback page shows only real activity (detections, restores, prunes). Windows client 2.5.93 already stopped reporting them; this clears the old entries from view.
# AEAntivirus 2.5.97
Version 2.5.97
- Linux ML parity (client 1.22.0): The Linux daemon now runs a 40-signal weighted ML engine ported from Windows — ELF anomaly detection (RWX segments, stripped section headers, unusual interpreters, packers), suspicious content signals (reverse shells, download-and-execute chains, base64 execution, credential access, persistence, cryptominers, ransomware, privilege escalation), script analysis (suspicious shebangs, PHP webshells, encoded PowerShell), and network/exfiltration signals (C2 strings, hardcoded IPs, DGA domains, Telegram/Discord webhooks). Score ≥ 65 quarantines. Also checks the fleet threat blocklist before heuristics.
- Android ML parity (client 1.24.0): New on-device URL ML (12 phishing signals) wired into QR code scanning — phishing links score 0-100 with contributing signals shown. New APK static analysis scores installed apps on dangerous permission combos (SMS spyware, admin+overlay ransomware patterns), native code loading, obfuscated package names, and known-malware patterns. APK hashes checked against the fleet blocklist.
- Smarter training: The server now counts verdicts since the last model training. At 50 new verdicts, the admin ML panel flags "retraining recommended" with a one-click Train now button, and Windows clients (2.5.92+) show a tray nudge if the model hasn't been trained in 7 days. New endpoints: GET /api/ml/training-status, POST /api/ml/trained.
- Update check dialog fix (client 2.5.92): The status dialog now pumps messages during the blocking HTTP call so it paints reliably instead of appearing blank.
# AEAntivirus 2.5.96
Version 2.5.96
- Fleet threat sharing (client 2.5.91): When you confirm a threat on one device, its SHA256 is shared instantly with every other device on your license. Those devices block the file on sight — no waiting for signature updates. See Security Operations → Fleet threats in the admin panel.
- Phishing URL protection (client 2.5.91): New 12-signal ML model scores URLs for phishing risk: IP-based URLs, suspicious TLDs, punycode/homographs, URL shorteners, excessive subdomains, phishing keywords, @-symbol credential phishing, hyphen-heavy domains, non-standard ports, missing HTTPS on login pages, and young-domain heuristics. URLs scoring over 70/100 are blocked by Web Protection.
# AEAntivirus 2.5.95
Version 2.5.95 (admin-only)
- Fixed admin panel JS syntax error: the Forensic sandbox manual-queue Clean/Malicious buttons used an incorrectly escaped quote in the served inline script, which broke parsing of the entire admin script block (blank dashboard). The buttons now emit correctly escaped quotes.
Version 2.5.94
- Trained ML model (client 2.5.88): Gradient-boosted decision trees trained on your verdicts — learns signal *interactions* (e.g. packed + network activity scores high even when each alone is weak). Click Train model in Security Operations → Detections → Trained ML model; see accuracy/precision/recall on a holdout set. Clients download the model and blend it 50/50 with the weighted score. Needs 20+ verdicts (50+ recommended).
- Sandbox queue (client 2.5.88): New Sandbox button in Quarantine — upload a suspicious file (with your confirmation) for manual analysis. Download it from Security Operations → Forensic sandbox → Manual analysis queue, run it in your own VM, mark it Clean or Malicious. The verdict flows back to the client and into ML training data. *Phase 1 is manual; automated detonation is a future phase.*
Version 2.5.93
- Network ML (client 2.5.87): 6 new signals watch per-process network behavior over 10-minute windows — C2 beaconing (periodic connections with low interval variance), data exfiltration (>10MB to non-allowlisted hosts), DGA domains (high-entropy DNS queries), port scans, Tor exit nodes, and suspicious ports.
- API sequence analysis (client 2.5.87): 5 new signals detect malicious API call chains — injection (VirtualAllocEx → WriteProcessMemory → CreateRemoteThread), persistence chains, crypto-ransomware sequences, keylogger patterns, and exfiltration chains. Only armed processes are tracked (no overhead on clean processes).
- Cloud Brain (client 2.5.87): The client now sends lightweight file features (~2KB JSON, never the file) to POST /api/cloud/score for an extended-model score combining community reputation, server-tuned weights, and your verdict history. Runs async (500ms timeout, never blocks scans); results are cached by SHA256 and blended 60/40 with the local score.
- 116 total signals (was 105).
Version 2.5.92
- 105-signal AI detection (client 2.5.86): The heuristic engine now checks 105 signals (was 60) — 40 new static signals (.NET obfuscation, stolen-cert patterns, APC/thread-hijack injection, doppelganging, bootkits, cryptominers, keyloggers, anti-analysis, document exploits including CVE-2017-11882 and CVE-2017-0199) plus 5 behavioral signals.
- Behavioral engine: New 5-minute sliding-window behavior tracking. Watches what processes DO — mass file encryption, shadow-copy deletion, persistence chains, C2 beaconing, process hollowing — and adds behavioral signals to the detection score in real time.
- Real-time learning: Every Confirm threat / False positive verdict now immediately nudges the firing signals' weights (TP +1, FP -2). No waiting for the weekly recompute. The admin panel shows "Learning applied" confirmation after each verdict.
- Auto-threshold guardrails: The system monitors your last 50 verdicts and auto-adjusts the detection threshold — FP rate over 5% raises it (less aggressive), zero false positives lowers it (more aggressive). Every adjustment is logged and visible in the admin ML page. Disable anytime with the Auto-adjust toggle.
Version 2.5.91 (admin-only)
- Verdict endpoint now handles missing detection rows: Marking a quarantine item as false positive no longer fails with "Reported 0 item(s) as false positive, 72 failed" when the detection row is missing (already resolved, or path case mismatch). The endpoint now tries an exact path match, then a case-insensitive match, and if still not found, still applies the false-positive exclusion and queues the quarantine restore — the user's intent (never flag this file again) is honored regardless.
Version 2.5.90
- 60-signal AI detection (client 2.5.80): The heuristic engine now checks 60 signals (was 45) — scheduled-task persistence, process injection (CreateRemoteThread+WriteProcessMemory), credential access (lsass/SAM), shadow-copy deletion, DGA domains, C2 beaconing, Office macro auto-execution, PowerShell in macros, LNK abuse, living-off-land binaries (certutil, bitsadmin, mshta with suspicious args), WMI event subscriptions, service installs.
- Smarter learning: Verdicts now decay with a 90-day half-life — recent decisions weigh more than old ones. The false-positive RATE (not just count) drives weight adjustments. Signals with fewer than 5 verdicts show "low confidence" in the admin panel.
- Detection explainability: Every AI detection stores its score (0-100) and per-signal score contributions. Click "why?" on any detection in the admin panel to see exactly which signals fired and how many points each added.
- Threshold advisor: New card in Security Operations → Detections analyzes your verdict history and recommends the optimal detection threshold, with one-click apply. The threshold is now server-configured (clients download it with the weights).
- Status page honesty: The client Status page green "You're protected" now requires ALL core protection on (Real-time, Firewall, Web Protection, Advanced RTP). If any are off, it shows "Attention needed" instead of a misleading green.
Version 2.5.89 (admin-only)
- Revenue dashboard — new Revenue section under Customers: total revenue, this month, MRR, active subscriptions, revenue by product, and recent transactions (live from Stripe).
- Per-customer billing — Billing button on each account row shows subscriptions, invoice history, payment methods, and lifetime total.
Version 2.5.88 (admin-only)
- AlienVault OTX via REST API: OTX's TAXII 2.1 server 404s on its objects endpoint (broken on their end), so the OTX feed now pulls through the OTX REST API v1 instead — subscribed pulses, then each pulse's indicators (MD5/SHA-1/SHA-256 hashes, IPv4, domains, URLs). SHA-256 hashes go straight into the signature database like ThreatFox/MalwareBazaar. No collection ID needed anymore — just paste the API key in Protection → AlienVault OTX and enable it.
Version 2.5.87 (admin-only)
- Verdict resolution workflow: Confirming a threat or marking a false positive now REMOVES the detection from the list (previously it stayed). False positives also queue a quarantine restore — the file is restored to its original location automatically. Confirmed threats are shared to the new Community threats section (Security Operations) showing threat name, hash, report count, and timestamps — the shared-protection list Austin asked for.
- Client verdict endpoint: New
POST /api/detections/verdict (install-authenticated) lets the Windows client report FP/confirm directly. Client 2.5.75 adds "Report false positive" and "Confirm threat" buttons to the Quarantine dialog.
Version 2.5.86 (admin-only)
- MISP feed buttons report errors instead of failing silently: The Pull/Edit/Delete buttons in the MISP/STIX feed table could fail with no feedback at all.
mispPull() now shows an alert and a message-line entry with the HTTP status and server error text on any failure, so the real problem (bad key, wrong collection id, unreachable host) is visible immediately instead of a dead click.
Version 2.5.85 (admin-only)
- ML feedback loop expanded (45 signals): The heuristic AI now tracks 45 detection signals (up from 22), including packer detection (UPX/AsPack/FSG), PE anomalies (no imports, suspicious entry point, stripped relocations), persistence clues (registry/WMI imports), encoded payloads (PowerShell, base64 blobs, embedded PEs), and certificate issues. Your Confirm-threat / Mark-false-positive verdicts auto-tune each signal's weight — signals you confirm get stronger, false positives get weaker. New Reset learning button clears the tallies back to base weights. Client 2.5.74 implements the new signals.
- Fixed ML tag mismatches: Two signal tags (
high-entropy, crypto-file-enum) were emitted with non-canonical names so verdicts never credited them. Now fixed — all 45 signals use exact canonical tags.
Version 2.5.84 (admin-only)
- Dashboard tiles populate on first load: The admin dashboard's tiles (Installs, Licenses, Trials Expiring, etc.) used to show "—" until the page was refreshed. The cause: a 2.5.64
showTab() optimization that skipped loader execution when already on the target tab — and since the tab initializes as dashboard, the very first visit returned early without ever running the data loaders. Now the first tab show always runs the loaders; the skip only applies afterward.
Version 2.5.83 (admin-only)
- Fixed manifest version overwrite: The
/updates/manifest.json endpoint had logic that overwrote the manifest's client version with RELEASE_VERSION (the server version) when it judged the manifest "stale". This broke client updates — the manifest would claim version 2.5.82 (server) when the actual client build was 2.5.73, so the client downloaded the "update", restarted, and the version number never changed (infinite update loop). The manifest version is now ALWAYS read from disk and never overwritten; server and client versions are independent.
Version 2.5.82
- False-positive exclusions (client 2.5.73): When you mark a detection as a false positive in Security operations → Detections, the file path is now recorded per-device in a new
fp_exclusions table. The client downloads its exclusions from the new install-authenticated GET /api/fp-exclusions endpoint (same update cycle as the whitelist) and never flags that file again on that PC — no popup, no quarantine, no cloud report. Path matching is case-insensitive with normalized slashes. - Confirmed threats shared fleet-wide (client 2.5.73): When you mark a detection as "Confirm threat", the file's SHA-256 hash is now added to the
community_threats table (the client reports the hash with each detection; new sha256 column on detections). Other devices pick it up through the threat feed and reputation checks — the opposite of FP exclusions.
Version 2.5.81 (admin-only)
- Mark any detection as false positive: The Security operations → Detections table now shows "Confirm threat" / "Mark false positive" buttons on ALL rows, not just detections with ML signal tags. The
/api/admin/detections/:id/verdict endpoint now records the verdict directly on the detection (new verdict column) even when no ML signal tags are present, instead of returning 400.
Version 2.5.80 (admin-only)
- Banner clarity: The threats-stopped ticker banner now reads "has stopped X threats all-time and counting" instead of "has stopped X threats and counting", making clear it's the lifetime total (vs the dashboard's 30-day detections card).
Version 2.5.79 (admin-only)
- Threat feed table readability: The Protection → Threat feed tables (Blocked domains #t5, Blocked IPs #t6) were hard to read — long domains/IPs wrapped mid-string and cells were cramped. Domain/IP cells now use a monospace font with
white-space: nowrap and ellipsis overflow (max 280px), all table cells have increased padding (8px 10px) and 13px font size.
Version 2.5.78
- Remote support billing notice (client 2.5.70): When a remote session starts on the customer's PC, a popup now explains the billing: 3 free sessions per calendar month, then $75 per session.
- Admin waive button: The Fleet → Remote support session history table now has a "Waive fee" button on billed sessions, so the $75 charge can be waived with an optional reason.
- Fixed false "outdated" badge: The version comparison stripped
-win/-srv/-android/-linux suffixes but not -WC/-LX, so clients on the latest version showed "outdated". The strip regex now covers -wc and -lx (case-insensitive). - Fleet installs table fits without horizontal scrollbar: 11 less-critical columns (Mode, Sigs, Scanned, Lifetime, Score, Plan, Expires, Last full, Last quick, Tags, Next scan) are now hidden via CSS. Kept visible: Hostname, Customer, Platform, Version, Heartbeat, Detected, Modules, License, Remote actions. Hidden data remains available in the install detail view.
Version 2.5.77 (admin-only)
- Re-added the 24 admin frontend loaders from 2.5.73 (reverted in 2.5.75 after a misdiagnosis): Resellers (loadResellers, rsSave, rsRevenue, loadPayouts, payoutMarkPaid, payoutCancel), Dunning (loadDunning), MISP (loadMisp, mispSave, mispCancelEdit, loadMispRuns, mispPull, mispEdit, mispDelete, mispKindChange), Trust signers (loadTrustSigners, tsAdd, tsRemove), Install map (loadInstallMap, geoSave), Remote (loadRemote), Retention (loadRetention), File cleanup (loadFileCleanup, saveFileCleanup). The functions were extracted byte-identical from the 2.5.73 commit and inserted at the same location; the real 2.5.76 blank-page bug was a stray
async keyword from the Remote-button insertion, now fixed.
Version 2.5.76 (admin-only)
- Fixed blank admin panel: a stray
async keyword was left on its own line before async function startRemoteSession (leftover from inserting it ahead of async function sendCmd), causing "Uncaught ReferenceError: async is not defined" at page load which killed the entire admin script. The stray keyword is removed; the admin panel loads normally again.
Version 2.5.75 (admin-only)
- Reverted the 24 frontend functions added in 2.5.73 (Resellers, Dunning, MISP, Trust signers, Install map, Remote, Retention, File cleanup loaders). They caused "Uncaught ReferenceError: async is not defined" at admin page load, leaving the entire admin panel blank. The functions will be re-implemented in smaller, tested batches. Kept from 2.5.74: the "🖥 Remote" button on Fleet install rows, startRemoteSession(), the sendCmd async fix, and the two onclick rename fixes (runFileCleanup, runRetentionPurge).
Version 2.5.74 (admin-only)
- Added "🖥 Remote" button to each Fleet install row (next to Update/Reboot/Uninstall/Lock/Unlock). Starts a remote support session via POST /api/admin/remote/session/start and opens the viewer in a new tab. The backend already existed; the UI trigger was missing.
Version 2.5.73 (admin-only)
- Restored 24 missing admin frontend functions (Austin: "Build all"): Resellers (loadResellers, rsSave, rsRevenue, loadPayouts + payoutMarkPaid/payoutCancel), Dunning (loadDunning), Install map (loadInstallMap, geoSave), Remote (loadRemote), MISP (loadMisp, mispSave, mispCancelEdit, loadMispRuns + mispPull/mispEdit/mispDelete/mispKindChange), Trust signers (loadTrustSigners, tsAdd, tsRemove), Retention (loadRetention), File cleanup (loadFileCleanup, saveFileCleanup). All backends already existed; only the UI loaders were missing.
Version 2.5.72 (admin-only)
- Confirm threat / Mark false positive buttons now show visible feedback: buttons disable on click, clicked button shows "✓ Recorded" on success.
Version 2.5.71 (admin-only)
- "Confirm threat" / "Mark false positive" buttons now always give feedback:
mlVerdict() previously had no try/catch, so if the fetch failed or the response wasn't JSON it exited silently — the buttons appeared to do nothing. Network errors now show an alert (API errors already showed alerts).
Version 2.5.70 (admin-only)
- All admin tabs now stay live: the 60-second poll previously refreshed only the active tab's data, so other tabs went stale until visited. The poll now refreshes every tab's already-loaded data (only loaders marked in
LOADED_FNS run — unvisited lazy sections are not triggered). The INP-aware skip (no poll if the user interacted within the last 5 seconds) still protects interaction performance.
Version 2.5.69
- Dashboard data now stays live: the 60-second poll was skipping the shared
load() function, so dashboard cards (installs, accounts, licenses, detections, etc.) went stale after the initial load and only updated on a manual page refresh. The poll now includes load() — data stays fresh automatically. The INP-aware skip (no poll if the user interacted within the last 5 seconds) and the staggered/tab-aware fetching still protect interaction performance.
- Add-on overlap cleanup: renamed and clarified overlapping add-ons — Password Vault family sharing (was "Family password vault"), Scam-call blocker (Android) platform label, PC tune-up vs System Optimizer differentiation, Dark web monitoring vs free Dark Web Scanner distinction, plus clearer copy for Parental insights, Second-opinion scanner, and Homework mode. Identity-protection add-ons now grouped under their own section on the account page.
Version 2.5.68
- Fixed admin logout redirecting to the wrong login page:
/api/admin/logout was redirecting to /login (customer login) after a form POST. Now redirects to /admin-login (the admin sign-in page), so logging out of the admin console no longer lands on the customer login form.
- 3 new paid add-ons: WiFi Security Scanner (automated WiFi encryption + device audit), Scam-Call Blocker for Windows (server blocklist synced to Scam Defense page), and Child Safe Search (DNS-enforced SafeSearch on Google/Bing/DuckDuckGo/YouTube, works with Parental Controls).
- Overlap audit: published ADDON_OVERLAP_AUDIT.md; merged password vault products in docs.
Version 2.5.67
- Fixed admin logout still leaving the user logged in:
adminLogout() now submits a form POST instead of fetch (browsers process Set-Cookie on full form navigations more reliably than on fetch responses), and /api/admin/logout now redirects to /login for browser form POSTs (returns JSON only for API clients that send an Accept: application/json header).
- System Optimizer: New sidebar page with three tools. Junk cleanup scans temp files, browser caches, Recycle Bin, thumbnail cache, Windows Update downloads, and log files — shows reclaimable space and confirms before deleting. Startup manager lists programs that launch at boot (registry Run keys + Startup folders) with one-click enable/disable. Registry cleaner finds orphaned uninstall entries and dead file associations, creates a .reg backup automatically before removing anything.
Version 2.5.66
- Fixed admin logout not actually logging out:
adminLogout()'s fetch was missing credentials: 'same-origin', so the browser never sent the ae_admin cookie — the server couldn't identify and delete the session, and the user stayed logged in after clicking "Log out". Now the fetch sends credentials and the session is properly invalidated server-side.
- Password Manager requires Password Vault add-on: The client's password manager page now checks for the paid vault subscription. Without it, users see a prompt to purchase from their account page.
Version 2.5.65
- Fixed the admin "Log out" button doing nothing: it called
adminLogout(event) but the function was never defined (ReferenceError). Added adminLogout() which POSTs to /api/admin/logout and redirects to /login.
- Password Manager (built-in): Encrypted password vault on your PC — AES-256 encryption, master password with PBKDF2 key derivation (100,000 iterations). Add/edit/delete entries (site, username, password, notes), password generator, one-click copy with 30-second clipboard auto-clear, auto-lock when idle, optional zero-knowledge cloud sync of the encrypted vault.
Version 2.5.64
- From Austin's profiler trace:
history.replaceState was taking 277ms per section click. Now skips replaceState when the hash is already correct (no redundant calls). showTab() now returns early when clicking the already-active tab (avoids rebuilding the subnav and the resulting Layout/style-recalc work — 390ms Layout + 190ms style recalc hotspot).
- Parental Controls: Web filtering by category (adult, gambling, social, gaming, streaming, dating) enforced at the DNS layer, per-day internet time limits, and an activity log of allowed/blocked web requests. Manage on the PC (PIN-gated Parental Controls page) or fleet-wide from the admin panel; activity uploads to the admin panel for review.
Version 2.5.63
- Fixed hung dashboard cards and empty Customers table: the
safeFetch in load() had no timeout — if any of the 8 admin API endpoints hung, Promise.allSettled never resolved and load() never completed, leaving cards (INSTALLS, ACCOUNTS, LICENSES) stuck at "—" forever. Now a 15-second AbortController timeout on each fetch; hung endpoints abort and load() continues with partial data.
- How to Use guide: The Feature Guide dialog (What's New page) now opens with a comprehensive How to Use section covering Real-time protection, Scans, Firewall, Web protection, Quarantine, Ransomware protection, Scheduled scans, Exclusions, Cloud lookup, DDoS Protection, and VPN.
Version 2.5.62
- Fixed the #1 admin performance hotspot:
loadThreats() was rendering ALL threat-intel IOCs (domains + IPs) with no limit — profiling showed it taking 4.8 seconds (43.7% of main-thread time). Threat feeds can have tens of thousands of entries. Now limited to 100 rows per table (domains, IPs) with a "Showing 100 of N" notice, matching the pagination pattern used elsewhere.
- Expanded threat detection: 25 new LOLBins (debuggers, credential theft tools, recon utilities), 18 new suspicious command patterns (lateral movement, fileless execution, Defender evasion), and 3 new AI scoring signals for credential theft, lateral movement, and fileless attacks.
Version 2.5.61 (admin-only)
- O(1) section switching:
showSec() no longer runs querySelectorAll over all instantiated tabs' sections on every click — now removes 'on' from the single visible section and adds it to the target by ID. Helps the Protection tab (16 sections).
Version 2.5.60
- Fixed MISP threat-intel scheduler crash: "ind.hashes is not iterable" (visible in server logs). All three indicator loops (
ind.hashes, ind.urls, ind.certs) are now guarded with Array.isArray before iterating.
- DDoS Protection add-on (paid): Protect your network from DDoS attacks — automatic SYN/UDP/ICMP flood detection with attacker blocking via Windows Firewall (30 min blocks), plus botnet guard to prevent your PC from participating in outbound attacks. Runs as a background service thread, gated by the paid add-on. Manage from the admin panel (grant/revoke per account).
Version 2.5.59 (admin-only)
- Fixed dashboard cards not populating: the 2.5.57 staggered fetch deferred non-critical endpoints behind idle waits, leaving cards (installs, accounts, licenses) empty on initial load. Initial page load now fetches all 8 endpoints in parallel; staggering applies only to subsequent 60s poll refreshes.
Version 2.5.58 (admin-only)
- Staggered admin fetch (2.5.57) is now tab-aware: the active tab's endpoints are fetched immediately instead of deferred. Fixes Customers tab showing empty accounts.
Version 2.5.57 (admin-only)
- Fixed broken section selector (
#adm .sec → #adm-sections .sec) - Deferred
window.scrollTo via requestAnimationFrame on tab switches TAB_RERENDER now dirty-checks datasets (skips re-render if unchanged)- Staggered the 8 shared
load() fetches (critical first, rest via requestIdleCallback) - 60s poll now skips if user interacted in last 5s (INP-aware)
- Hoisted
ESC_MAP (was re-created on every esc() call)
Version 2.5.56 (admin-only)
Admin interaction responsiveness (INP) optimization:
- Section-lazy data loading: on tab switch, only the visible section's data loaders run immediately (Protection tab: 18 → ~3 requests). Other sections load when clicked, or are deferred via requestIdleCallback.
- Signatures table paginated at 100 rows (was building 500 rows × 8 regex ops in one shot).
- 60s poll now only refreshes already-loaded sections instead of triggering lazy loads.
Version 2.5.55 (admin-only)
- Admin console fixes:
setT() is now a global null-safe text setter (was local to load()), fixing ReferenceError: setT is not defined in loadThreats. loadPresets() is now guarded for the lazy-loaded Presets tab (t7), fixing TypeError: Cannot set properties of null (setting 'innerHTML'). No customer-facing changes; the client stays at 2.5.69.
Version 2.5.54 (admin-only)
- Admin tab-switching speed: removed the 2.5.46 behavior that dropped a tab's table DOM on every tab switch. With 2.5.49 lazy-loading, initial memory is already bounded, and rebuilding tables on every switch caused slowdowns when clicking back and forth. Visited tabs now stay live. No customer-facing changes; the client stays at 2.5.69.
Version 2.5.53 (admin-only)
- Admin console fixes: More null-safety for lazy-loaded tabs —
c-susp, c-bugs, and th-meta counters now use the null-safe setT() helper, and the t3/t4/t5/t6 table renders are guarded against non-instantiated tabs. Fixes TypeError: Cannot set properties of null (setting 'textContent') on admin load.
Version 2.5.52 (admin-only)
- Admin bugfix: fixed
Uncaught (in promise) TypeError: Cannot set properties of null (setting 'innerHTML') at renderAccounts — all five table renderers (renderAccounts, renderInstalls, renderDets, renderSusp, renderQuar) now return early when their table container doesn't exist (lazy-loaded tab not yet instantiated). Visiting the tab renders normally via showTab() → ensureTabDOM() → runTabLoaders(), so no functionality is lost. No customer-facing changes; the client stays at 2.5.69.
Version 2.5.51 (admin-only)
- Admin bugfix: fixed
Uncaught (in promise) TypeError: Cannot read properties of null (reading 'value') at renderInstalls — the table renderers (renderInstalls, renderDets, renderSusp, renderQuar) now tolerate missing filter elements (lazy-loaded tabs not yet instantiated) by using default empty filter values instead of throwing. No customer-facing changes; the client stays at 2.5.69.
Version 2.5.50 (admin-only)
- Admin bugfix: fixed
Uncaught TypeError: Cannot read properties of null (reading 'addEventListener') on /admin — table search inputs inside lazy-loaded templates now get their event listeners bound defensively via bindTableSearch(), re-run after each template instantiation in ensureTabDOM(). No customer-facing changes; the client stays at 2.5.69.
Version 2.5.49 (admin-only)
- Admin tab lazy-loading: inactive admin tabs now live in inert
<template> tags instead of the live DOM — only the active tab's sections are parsed into DOM nodes (down from ~4,800 to ~700 on load). The 248KB admin guide, orphaned dead sections, and per-tab scripts instantiate on first visit. Also surfaced two previously unreachable admin sections (AlienVault OTX, YARA + Consensus) in the Protection tab. No customer-facing changes; the client stays at 2.5.69.
Version 2.5.48 (admin-only)
- Admin API fixes: the
/api/admin/installs endpoint now caps results at 2000 rows (was unbounded) and no longer re-reads the remote-commands file from disk for every install row (one read instead of N). No customer-facing changes; the client stays at 2.5.69.
Version 2.5.47 (admin-only)
- Admin memory optimization: the Installs, Detections, Suspicious files, and Quarantine tables are now paginated (100 rows per page with Prev/Next controls) instead of rendering every row at once, and switching admin tabs now releases the previous tab's rendered table rows from memory. This fixes the multi-GB RAM growth seen on the admin page with large fleets. No customer-facing changes; the client stays at 2.5.69.
Version 2.5.41
- Settings window scrollbars restyled: the Settings popup's white scrollbars are replaced with the same slim dark-groove, glowing-cyan-thumb bars as the sidebar (both vertical and horizontal). Drag the thumb, click the groove to page, or use the mouse wheel.
Version 2.5.40
- Restyled sidebar scrollbar: the default white slider is gone — the sidebar now has a slim dark groove with a glowing cyan thumb (brightens on hover). Drag the thumb, click the groove to page, or use the mouse wheel, exactly as before.
Version 2.5.6
- Startup crash fix: disabled the three new 2.5.0 background monitors (cryptominer, stalkerware, scareware) — one of them crashes the process at ~16s with an access violation. They'll be re-enabled one by one once the culprit is found.
Version 2.5.5
- Startup crash fix: disabled the automatic initial quick scan — it crashed (0xC0000005) on certain installer executables (AEAntivirus-Setup, ChromeSetup). Real-time protection remains fully active. Manual scans can be run from the UI.
Version 2.5.4
- Startup crash fix: the scanner crashed with an access violation when scanning our own installer (AEAntivirus-Setup*.exe). The scanner now skips our own installer files — they're our own signed binaries and don't need scanning.
Version 2.5.3
- Scan crash diagnostics: the initial quick scan was crashing the UI at ~16s. Added per-file crash tracking so the log identifies the exact file causing the crash.
Version 2.5.2
- Startup crash fix: the integrity monitor's alert notification could crash the UI when fired during startup (direct PostMessage raced UI initialization). Alerts now use the file-based notify channel only — the UI picks them up safely via its normal poll. Also fixed the redundant immediate periodic integrity check after the startup check.
Version 2.5.1
- Startup crash fix: 2.5.0 could close seconds after opening. Background monitors (integrity, cryptominer, stalkerware, scareware) now wait 15 seconds for startup to settle before their first sweep, and crash logs now name the active monitor plus thread ID for diagnosis.
Version 2.4.1
- Heuristic Tuning page layout fix: the description text, column headers, and Refresh button no longer overlap each other.
- Full-width site: aeantivirus.com now uses the whole browser window on any monitor size — no more narrow centered column.
Version 2.1.3
- Fixed hidden description text: the Wi-Fi and ARP detector listboxes were overlapping their description labels, hiding the first line of text. The labels are now positioned below the listboxes.
Version 2.1.2
- What's New stays on top: the What's New / Feature Guide window now stays above the main UI until you close it.
Version 1.97.3
- AE logo on every popup: all Windows popups and dialogs — the firewall/program prompts, PIN dialogs, rule editors, AI chat windows, restore/vault/scan dialogs, and the installer wizard — now show the AE logo in their title bars instead of the generic Windows icon. Every AE window class registers the embedded app icon.
- Build fix: the 1.95.0 ML-behavior cross-check referenced a file-integrity query that was never implemented, breaking the link. A stub returning "not flagged" is in place until the real query lands.
Version 1.95.0
- Detection engine round 2 — threat intel feeds: two more sources join the feed list — SSLBL (abuse.ch malicious SSL certificates, CC0 public domain) and PhishTank (verified phishing URLs; needs a free API key, so it's supported but not a default). Feeds now score themselves: the server tracks what fraction of each feed's hits were true positives vs. noise, and a feed whose precision drops below 50% (10+ samples) is auto-paused — the admin sees the precision and can re-enable with one click. Also fixed two SSRF-filter bypasses (trailing-dot hostnames, IPv4-mapped IPv6) found while testing.
- Detection engine round 2 — cloud network: the cloud now answers file-hash (SHA-256) reputation lookups in addition to URL/host/IP — only the hash is ever sent, never file contents. Failover: the client tries the primary cloud endpoint, then a secondary on timeout/errors with backoff (fleet policy can set the secondary URL). First-seen tracking: cloud verdicts carry a first-seen timestamp; the UI shows "first seen X days ago". Contributor weighting: reports from long-lived, low-noise installs weigh more in the community corpus (still needs 2+ distinct installs; weight only breaks ties toward suspicious).
- Detection engine round 2 — heuristics: the static scorer adds .NET assembly checks (suspicious attributes, mixed-mode anomalies), resource-section analysis (version-info mismatches like product name != internal name), and section-entropy outlier detection. The behavior scorer adds C2 beaconing detection (periodic connections to the same host with regular intervals), process-hollowing signals (suspended start + unmapped memory writes), and integrity-monitor cross-checks. A new trusted-signer allowlist (Microsoft, Google, Mozilla, Apple, major AV vendors — admin-viewable/editable, published to clients via the update channel) reduces heuristic weight by 30 points for well-known vendor signatures, cutting false positives.
- Unified threat score: cloud verdict + heuristics + feed hits now combine into one documented scoring rule (extends the 1.89.0 ensemble; heuristic hits stay "Suspicious" advisories, never ML claims, never definitive verdicts).
- Cleaner homepage: paid add-on cards in "Everything included" no longer show prices (prices live on the account/checkout pages); the long stack of "New in" sections is consolidated — only the newest shows in full, older ones sit behind a "Show older updates" expander. The in-app What's New page was reformatted with shorter lines and clearer bullet spacing.
Version 1.94.0
- 6 new paid add-ons: Trades pack ($3.99/mo) — job-site device theft protection + invoice-scam detection for contractors. Auto shop pack ($4.99/mo) — shop PC hardening, customer data handling, key-programmer tool protection. Field crew pack ($2.99/mo) — theft alerts and rugged-use settings for landscaping/field crews. Professional pack ($9.99/mo) — privilege-grade baseline for lawyers, accountants, and clinics with client data. Content takedown service ($9/takedown, bookable) — DMCA + non-consensual content removal assistance. Nanny cam security check ($19 one-time, bookable) — verify nursery/child cameras aren't internet-exposed.
Version 1.93.0
- 6 new paid add-ons: Ransomware warranty ($9.99/mo) — money-back guarantee if ransomware encrypts your files while required protections are on; full terms (required settings, 30-day claim window, log verification, exclusions) are shown on the card before you buy, and your acceptance is recorded. Small-biz SOC lite ($29/mo) — a human reviews your flagged alerts and sends plain-language verdicts. Pen-test lite ($99 one-time, bookable) — external + internal test with a human-reviewed findings report. Homeschool pack ($3.99/mo) — kid-safe defaults, content filtering, screen-time bundle. Storm prep pack ($1.99/mo) — backup checklist, offline comms plan, device prep guide for hurricane season. Medicare scam protection ($1.99/mo) — senior-focused medical scam call/message blocking + education.
Version 1.92.0
- New "System Care" page: 5 PC-maintenance tools. Proxy detector watches your system proxy setting and alerts if a proxy gets enabled or changed (malware loves silent proxies). Hotspot guard checks your Windows mobile hotspot is using strong WPA2/WPA3 security when sharing. Old download cleaner finds files in Downloads older than N days (you pick N, 7 minimum) and moves them to the Recycle Bin after a confirm. Context menu cleaner lists third-party right-click handlers, lets you disable the junk ones, and restores them anytime. Store app audit inventories your installed Store/sideloaded apps and flags non-Microsoft publishers.
- Bonus: QR safety scanner (Android): new in the Android app — point the camera at a QR code, and it shows the expanded URL plus a safety verdict BEFORE you open it. No more blind taps on mystery QR codes.
Version 1.91.0
- New "Kernel Guard" page: 7 deep-system defenses. Browser profile stealer guard watches Chrome/Edge/Firefox credential stores and flags unusual processes when they change. Shadow copy guardian kills vssadmin/wmic shadow-deletion attempts (the classic ransomware step). LSASS protection (PPL) toggle, PowerShell Constrained Language toggle, AMSI/ETW tampering detector (alerts on security-setting tampering), ISO mount guard (alerts on new disc/image volumes), and honeytoken credentials (plants decoy files and alerts when touched).
Version 1.89.0
- Threat intel feeds ship ON by default: the server now pulls the URLhaus malicious-URL list and the Abuse.ch ThreatFox IOC feed automatically (both free for this use with attribution, shown in Admin → Threat intel). Admins can disable or replace them; the existing feed health monitoring covers the new sources.
- Cloud URL + IP reputation: the client can now ask the cloud about a website host or an IP address, following the same hash-only privacy pattern as file lookups (only the bare host/IP is sent — never file contents, paths, or full URLs with query strings). Web Protection and the firewall consult it.
- Smarter cloud caching: clean verdicts keep the 24h cache; suspicious/malicious verdicts re-query after 1h and unknowns after 30min, so bad verdicts self-correct fast and fresh intel is picked up quickly.
- Community threat sharing feeds the cloud corpus: opted-in installs' detections now count toward the server-side reputation (2+ independent reports marks a hash suspicious) — one user's detection protects everyone faster. Opt-in only, never automatic.
- Cloud status indicator: the Windows app's Status page now shows a live cloud threat-intelligence status line (connected/unreachable, last successful lookup).
- Heuristic engine upgrades (still weighted heuristics, not ML — "Suspicious" advisory language kept): the static scorer now checks file imphash fingerprints against a known-bad list shipped via the updater, flags unsigned binaries sitting in Windows system directories (Authenticode), and adds writable-code-section and packed-slack-section signals. The behavior scorer adds LOLBin detection (powershell/cmd/script hosts), parent/child anomaly detection (Office or a browser spawning a shell), and suspicious command-line patterns (encoded payloads, download cradles) via best-effort WMI.
- Cloud + heuristic ensemble: the documented combination rule — cloud CLEAN quiets borderline heuristic scores (40–59), cloud UNKNOWN + score 45+ still raises the advisory, score 60+ always flags.
Version 1.88.0
- New paid add-ons (6): Crypto trader pack ($4.99/mo), Military family pack ($2.99/mo), Expat pack ($4.99/mo), New-hire security onboarding ($15/hire), Landlord device turnover ($19/turnover), Kids' smartwatch safety ($1.99/mo per child).
- The two technician services (new-hire onboarding, landlord turnover) are bookable from your account — our technician contacts you within one business day to schedule. Every card states what happens after you buy.
Version 1.87.0
- New paid add-ons (6): SIM-swap guard ($1.99/mo), Home title fraud monitoring ($2.99/mo), Credit freeze manager ($1.99/mo), Deepfake call detector ($2.99/mo), Developer security pack ($4.99/mo), Holiday scam protection ($1.99/mo, Nov–Jan).
- Honest checkout: every add-on card now tells you plainly what happens after you buy — whether it activates instantly, needs a setup step from you, or is a scheduled technician service. The purchase confirmation page does the same.
Version 1.86.0
- Windows app: new "Network Watch" sidebar page: rogue DHCP server detector, Wi-Fi deauth attack watch, whole-LAN IoT default-password sweep, and an opt-in to donate idle CPU to community threat research (preference only for now; shared tasks arrive in a future update).
- Windows app: new "Scam Defense" sidebar page: a scam popup killer that auto-closes fake virus alerts / fake CAPTCHAs / tech-support scares, and an email header analyzer that traces the true origin with SPF/DKIM/DMARC verdicts.
- Windows app: Privacy Tools gains a built-in TOTP authenticator (secrets stay on your PC) and a DPAPI-encrypted secure note.
Version 1.85.0
- Windows app: new "System Guard" sidebar page with 8 features: USB charge-only mode (blocks USB storage data while charging), one-click Windows Sandbox file opener, Attack Surface Reduction rules toggle, Controlled Folder Access manager (toggle + protected-folder list), VBS + Memory Integrity status, realtime registry autorun monitor, open port audit with risk flags, and BitLocker recovery-key backup to your account.
Version 1.84.11
- Windows app: new "What's New" page in the sidebar. It opens by itself on first install and once after every update, and explains in detail what every feature does and how to use it (all sidebar pages, the 100 free + 100 paid add-ons, and tips). Reopen it any time from the sidebar.
Version 1.84.10
- Windows app: fixed the white ghost box left behind on the Net Hardening page (its listbox wasn't registered for page cleanup, same root cause as the Recall & USB fix in 1.84.8).
- Windows app: sidebar nav buttons no longer slide under the scrollbar — buttons stop before it, and long labels clip with an ellipsis instead of running underneath.
Version 1.84.3
- Windows app: sidebar nav now highlights the active page on all 25 pages (the 10 round-2 pages were missing). More prominent indicator: thicker cyan bar + cyan dot.
Version 1.84.2
- Windows app: tile buttons (Status page cards, Privacy tools, etc.) now use the same bright cyan border + glow + dark gradient styling as all other buttons.
Version 1.84.1
- Server: fixed update downloads for private GitHub repo. The manifest now points clients to
/updates/proxy/ URLs — the server fetches binaries from GitHub with API auth and streams them to clients, instead of raw.githubusercontent.com links that 404 on private repos. The download page uses the same proxy.
Version 1.84.0
- Admin: Audit Log. Every admin change to add-on config (enable/disable/hide/settings), pricing overrides, custom policies, and fleet policy is logged with timestamp, admin email, target, field, old and new values. New admin UI section to view and search the log.
Version 1.83.0
- Admin: Custom Policy Builder. Create named policies with chosen protection settings, set one as the default for all installs (per-install/group/user overrides still apply). Includes create/edit/delete/duplicate and clear default-policy indication.
Version 1.82.0
- Paid add-ons: 5 new "Identity & Estate" add-ons (see catalog).
Version 1.81.0
- Paid add-ons: 5 new "Seasonal & Pro" add-ons (see catalog).
Version 1.80.0
- Paid add-ons: 5 new "Life Events" add-ons (see catalog).
Version 1.79.0
- Paid add-ons: 5 new "Backup Packs" add-ons (see catalog).
Version 1.78.0
- Paid add-ons: 5 new "Business Security" add-ons (see catalog).
Version 1.77.0
- Paid add-ons: 5 new "Family & Business" add-ons (see catalog).
Version 1.76.0
- Paid add-ons: 5 new "Privacy Audits" add-ons (see catalog).
Version 1.75.0
- Paid add-ons: 5 new "Connectivity" add-ons (see catalog).
Version 1.74.0
- Paid add-ons: 5 new "Home Services" add-ons (see catalog).
Version 1.73.0
- Paid add-ons: 5 new "Dark Web & Services" add-ons (see catalog).
Version 1.72.0
- Windows app: new "Learn & Test" page with 5 free features: Proximity auto-lock, Lid-close lock, Tip of the day, Threat encyclopedia, Password strength tester.
Version 1.71.0
- Windows app: new "Focus & Safety" page with 5 free features: Focus mode, Presentation mode, Break reminders, Panic button, Guest mode.
Version 1.70.0
- Windows app: new "Net Tools" page with 5 free features: Router default-credential checker, Local network map, Wake-on-LAN sender, Metered-connection helper, Captive portal safety check.
Version 1.69.0
- Windows app: new "Net Monitor" page with 5 free features: Per-app bandwidth monitor, Data exfiltration alert, Firewall log viewer, Connection history, Wi-Fi password auditor.
Version 1.68.0
- Windows app: new "Disk Tools" page with 5 free features: Large file finder, Duplicate file finder, Secure file shredder, Folder locker, Startup delayer.
Version 1.67.0
- Windows app: new "Device Control" page with 5 free features: Camera kill switch, Windows debloater, Restore point guardian, Driver backup, Broken shortcut cleaner.
Version 1.66.0
- Windows app: new "Net Privacy" page with 5 free features: WebRTC leak toggle, Encrypted DNS (DoH) toggle, UPnP risk auditor, Bluetooth stranger alerts, Mic mute hotkey.
Version 1.65.0
- Windows app: new "Privacy Tools" page with 5 free features: Idle-time scanner, SHA256 hash calculator, Advertising ID disabler, Clipboard history wiper, Activity timeline clearer.
Version 1.64.0
- Windows app: new "Net Hardening" page with 5 free features: SMBv1 auto-disabler, LLMNR/NBT-NS spoofing guard, ARP spoofing detector, RDP inbound blocker, PDF JavaScript sanitizer.
Version 1.63.0
- Windows app: new "Recall & USB" page with 5 free features: Windows Recall disabler, USB auto-scan toggle, evil-twin Wi-Fi detector, local backup scheduler, Office macro stripper.
Version 1.62.0
- Windows app: fixed invisible button text on 12 content pages (Advanced, Family, Power User, System Tools, Smart Scan, Network, Privacy, Hardening, USB Defense, Security Hub, Cloud Storage, Performance). The WM_DRAWITEM fallback's class-name check was silently failing for content-page buttons; it now draws all unhandled owner-drawn controls with the standard themed painter.
Version 1.61.9
- New logo in the Windows app header: the main window now shows the new AE logo (embedded PNG, drawn with GDI+) instead of the drawn hexagon mark — the runtime-built tray icon uses it too.
- Scan-complete video removed: the spinning VERIFIED celebration video has been taken out of the Windows app, the Android app (1.23.3), and the homepage;
updates/scan-complete.mp4 is gone from the server. Scans finish with the standard completion notice again.
Version 1.61.8
- Scan-complete celebration: every finished scan now ends with the spinning AE logo "VERIFIED" animation — in a popup in the Windows app (MCI video, respects silent mode and quiet hours; the video downloads once from the server and is cached), in a dialog in the Android app (1.23.2), and embedded on the homepage. The video is served at
updates/scan-complete.mp4.
Version 1.61.7 (admin-only)
- Windows app: diagnostic build. Removed the 50-entry log cap so every button draw is logged. Open the Hardening page first after deleting the log to capture those button IDs.
Version 1.61.6
- New logo everywhere: brand-new AEAntivirus mark — the AE monogram with a checkmark for "verified" and circuit traces for the technology — now on the website (header, favicon, app icons), the Windows app and installer, and the Android app (launcher icon and welcome screen).
Version 1.61.5 (admin-only)
- Windows app: removed ETO_CLIPPED flag from button text drawing. If the clipping rectangle was misaligned on some systems, it would silently clip the text to nothing.
Version 1.61.4
- Windows app: update check and download page now report the correct version even when the server's local files are stale. The update manifest points to GitHub for fresh binaries.
Version 1.61.3 (admin-only)
- Windows app: expanded button diagnostics. The draw log now includes text position, size, button rect, text color, and font handle to pinpoint why text is invisible.
Version 1.61.2 (admin-only)
- Command Center: bulk hide/unhide. Select multiple add-ons with checkboxes (or select all) and hide/unhide them in one click.
Version 1.61.1 (admin-only)
- Command Center fix: resolved a JavaScript error (
BS is not defined) that prevented the add-on table from rendering.
Version 1.61.0
- Command Center: hide any add-on from all user interaction with one click. Hidden add-ons vanish from the catalog, account pages, checkout, and the Windows app until unhidden — stronger than disable, which only blocks new purchases.
- Device discovery: new Type column shows whether each device is on wireless or wired, detected from the reporting PC's network interface.
Version 1.60.9 (admin-only)
- Network monitor: duplicate devices no longer show up. If a device is reinstalled and gets a new ID, only its most recent report appears (matched by hostname).
Version 1.60.8 (admin-only)
- Network monitor: each device's report card now has a Clear button to remove it from the view. The device reappears on its next report (about a minute later).
Version 1.60.7 (admin-only)
- Revenue module: new Revenue section in the admin panel (Customers tab) showing total revenue, this month's revenue, MRR, active subscription count, revenue by product, and recent transactions — all pulled live from Stripe and filtered to AEAntivirus + add-on sales.
Version 1.60.6 (admin-only)
- Network monitor: devices can now be deleted from the admin panel (red X button per row). MAC addresses are normalized on insert (AA-BB-CC vs AA:BB:CC vs AABBCC all match) and existing duplicates are cleaned up on server start, keeping the most recently seen.
Version 1.60.5
- Trials: the 7-day trial countdown now starts when the license key is first used in an install, not when the key is created. Keys provisioned via page visit no longer lose trial days before install.
Version 1.60.4 (admin-only)
- Windows app: diagnostic build for the blank-button issue. Writes button draw diagnostics to %TEMP%\\ae_draw.log so we can see exactly why text isn't appearing.
Version 1.60.3
- Platform licenses: the Mobile plan now issues Android licenses. All other plans let you pick Windows or Linux at checkout, and the Custom plan offers all three platforms. Every installer still includes a 7-day free trial, platform-scoped automatically.
Version 1.60.2
- Subscribe page: removed the Android and Linux plans. Only the Windows plans (Personal, Family, Business, etc.) are offered now.
Version 1.60.1
- Windows app: rewrote button text drawing (ExtTextOutA with manual centering). The Hardening page buttons were rendering as blank shapes on some systems — text now draws reliably.
Version 1.60.0
- Add-on health checks (admin): the Command Center now shows a Health column for all 100 add-ons. Paid add-ons verify their backend (Stripe, SMTP, storage, telephony) — a missing dependency shows as a warning before a customer hits it. Free add-ons show when they were last seen running on a real PC (heartbeat data).
Version 1.59.1
- Windows app: fixed blank button text on the Hardening page (and any other affected pages). Button labels are now cached at creation and drawn from the cache — no more empty buttons.
Version 1.59.0
- Add-on Command Center (admin): one screen for all 100 add-ons — the paid 50 and the free 50. Toggle any add-on off to disable it globally (paid add-ons stop selling instantly: checkout blocked, hidden from the catalog; free features switch off on PCs at the next heartbeat via a client kill-switch). Edit per-add-on JSON settings, and monitor live usage: active purchases for paid, installs with the feature on for free.
- Windows app: 19 free features now honor the admin kill-switch (USB autorun protection, hosts guard, DNS watchdog, MBR guard, cam/mic alerts, travel mode, family-safe DNS, boot scan, download reputation, wake-for-scan, battery-aware scanning, context-menu scan, high-contrast, voice alerts, community sharing, clipboard guard, app lockdown, and more). The client also reports its feature states in the heartbeat for the Command Center monitor.
Version 1.58.5
- Windows app: every toggle button is now born with its label — no more blank buttons on USB Defense, Hardening, Privacy Watch, Network, Scanning, Power User, Family, Advanced, or Storage pages. Also fixed overlapping section headers on the Smart Scan and Advanced pages.
Version 1.58.4
- Windows app: the separate server status light and text are gone — the big status circle's checkmark now shows the server connection: green = connected, yellow = connecting, red = failed. Click the circle to retry the connection, hover it to see the connection detail. Protection state still drives the check-vs-exclamation shape.
- Windows app: the sidebar tab list now has a vertical scrollbar (plus mouse-wheel scrolling) so all 15 tabs stay reachable — the last tabs were getting cut off at the bottom of the window.
Version 1.58.3
- Windows app: fixed the invisible toggle buttons on USB Defense, Hardening, Privacy Watch, Network, Scanning, Power User, Family, Advanced, and Storage pages. Button labels (e.g. "Autorun protection: ON") now always paint.
Version 1.58.2
- Windows app UI fixes: all listboxes and text fields on the content pages now use the dark theme (no more white boxes on USB Defense, Hardening, Privacy Watch, or the incident timeline). Main window enlarged to 1360×920 with wider lists.
Version 1.58.1
- Add-on pricing editor (admin): change any of the 50 paid add-on prices from Admin → Paid add-ons. Prices take effect on the next checkout immediately — no code changes, no Stripe dashboard work.
Version 1.58.0
- Paid add-ons, batch 10 — the final paid 50 (Windows app + Android + server + My Account): Android TV guard (leanback launcher + sideload scan), second-opinion cloud scanner, scheduled family scans, grandparent help button, early-access club with feature voting.
Version 1.57.0
- Paid add-ons, batch 9 (Windows app + server + My Account): USB portable vault (AES-256 encrypted, opens on any PC), router hardening service, identity restoration cases, breach response hotline, social account guard.
Version 1.56.0
- Paid add-ons, batch 8 (Windows app + server + My Account): email mailbox protection (server-side phishing/malware filtering), travel bundle (public-Wi-Fi travel mode), gamer pack (fullscreen game mode), Photo organizer Pro (duplicate finder), file recovery (Recycle Bin restore scanner).
Version 1.55.0
- Paid add-ons, batch 7 (Windows app + server + My Account): extra seat packs (adds devices beyond your plan limit), server protection (hardened profile for Windows Server), secure-delete certificates (tamper-proof proof-of-erasure), custom block pages (your branding on threat warnings), incident response retainer (4-hour technician SLA).
Version 1.54.0
- Paid add-ons, batch 6 (Windows app + server + My Account): Wi-Fi intruder history Pro (every network your devices ever joined), crypto wallet guard (on-screen address verification + wallet app detection), digital inheritance vault (encrypted beneficiary release), offline update USB (quarterly mailed USB), white-glove onboarding (bookable technician setup).
Version 1.53.0
- Paid add-ons, batch 5 (Windows app + server + My Account): homework mode (blocks distracting apps during homework hours, in the Family page), smart-home device monitor (LAN devices reported to your account), SMS phishing filter API (Android), kids' profiles (safe Android mode), and parental insights Plus (weekly family security summary).
Version 1.52.0
- Paid add-ons, batch 4 (server + Windows app + My Account): external vulnerability scan (outside-in port probe with plain-language fixes), compliance checklists (HIPAA/PCI), ransomware rollback Plus (90-day version history in the app), extended threat history (1 year of detections), and the scam-call blocker list (Android app pulls it automatically).
Version 1.51.0
- Paid add-ons, batch 3 (server + My Account): website uptime monitor (10-minute checks, downtime email), domain & SSL monitor (daily cert-expiry checks with 30/14/7/1-day warnings), priority human support (front-of-queue badge), plus the new-PC setup and annual checkup technician services (bookable from My Account).
Version 1.50.0
- Paid add-ons, batch 2 (Windows app + server): family locator (device presence with arrival alerts), secure file sharing (encrypted expiring links), Driver Updater Pro (driver scan with restore-point safety), PC tune-up subscription (monthly automatic optimization), and the home network audit technician service (bookable from My Account).
Version 1.49.0
- Paid add-ons, batch 1 (Windows app + server): the 50-add-on catalog with one generic Stripe checkout (no per-add-on dashboard setup), technician service bookings (remote virus removal and more), automatic versioned cloud backup (scheduled, client engine + server), senior protection pack (simple mode + trusted-contact alerts), employee phishing training (campaigns, click tracking, coaching), and laptop anti-theft (locate + lock; remote wipe coming later). Admin gets a bookings queue, campaign overview, and grant/revoke for any catalog add-on.
Version 1.48.0
- Advanced Protection (Windows app + admin panel): a new sidebar page with a clipboard guard (blocks crypto-address hijacking), a cryptominer detector, app lockdown (default-deny for new executables), file integrity monitoring (critical system files), and threat screenshots for forensics. The admin panel shows each install's clipboard guard and app lockdown states. This completes the free 50 features.
Version 1.47.0
- Family & Learning (Windows app + server): a new sidebar page with a phishing drill (spot the phish, with explanations), a scam quiz (8 real-world scenarios), high-contrast mode toggle, voice alerts (spoken security warnings via SAPI), and opt-in community threat sharing (SHA-256 hashes only, never file contents). New
POST /api/community-threat endpoint.
Version 1.46.0
- Power User (Windows app + server): a new sidebar page with an Explorer context-menu toggle, quarantine cleanup (auto-delete old files, configurable retention), exclusion import/export, a CLI reference, and settings sync (upload/download via your account). New
POST /api/settings-sync and GET /api/settings-sync endpoints.
Version 1.45.0
- System Tools (Windows app): a new sidebar page with a startup manager (one-click enable/disable with safe restore), a browser extension audit (flags risky permissions in Chrome/Edge/Firefox), an app updater (winget-powered), disk health monitoring (SMART status), and a tracking-cookie cleaner.
Version 1.44.0
- Smart Scanning (Windows app + admin panel): a new sidebar page with boot-time scan (catches malware before login), download reputation scoring, one-click scan profiles (Quick, Full, Gaming, Work, Deep Clean), wake-for-scan (wakes a sleeping PC at 2am, scans, sleeps again), and battery-aware scanning (pauses when the battery is low). The admin panel shows each install's scan profile and wake-for-scan state.
Version 1.43.0
- Network Modes (Windows app + admin panel): a new sidebar page with public Wi-Fi security warnings, travel mode (maximum firewall strictness for a trip window), network profiles (Home/Office/Public with per-profile strictness), time-based firewall rules, and one-click family-safe DNS (blocks adult content). The admin panel shows each install's travel mode and family DNS states.
Version 1.42.0
- Privacy Watch (Windows app + admin panel): a new sidebar page with a spyware sweep (stalkerware, keyloggers, RATs, hidden remote tools), remote session alerts that scream when someone connects via RDP or a remote tool launches, a permission audit listing camera/mic/location access per app, camera/mic background-use alerts, and a social media privacy checkup guide. The admin panel shows each install's remote-alert and cam/mic-alert states.
Version 1.41.0
- Hardening (Windows app + admin panel): a new sidebar page with exploit protection (forces ASLR + DEP + anti-ROP mitigations onto vulnerable apps like browsers and Office), a hosts-file guard lock, a DNS change watchdog that alerts on silent DNS hijacks, an MBR/boot-sector guard that detects bootkit overwrites, and a Secure Boot / TPM / BitLocker dashboard with a guided enable wizard. The admin panel shows each install's hosts lock, DNS watchdog, and MBR guard states.
Version 1.40.0
- USB Defense (Windows app + admin panel): a new sidebar page with a one-click USB autorun protection (fully disables auto-run/auto-play exploits), a managed trusted-USB-device allowlist (trust connected sticks, remove old ones), a portable scanner creator (runs on PCs without AEAntivirus installed), offline signature export/import for air-gapped PCs via USB, and a rescue USB builder with instructions. The admin panel shows each install's autorun protection state.
Version 1.39.0
- Security Hub (Windows app + admin panel): a new sidebar page with a 0–100 security score (one-click "Fix issues" for whatever is dragging it down), an incident timeline of every detection/block/scan with plain-language "why was this blocked?" explanations, one-click CSV export of your threat history, and a weekly threat briefing digest. The admin panel shows each install's score in the fleet table and the full synced event timeline in the install detail view.
Version 1.38.10
- Check for Updates now shows it's working (Windows app): clicking the button immediately disables it, changes its label to "Checking...", and pops a "Checking for updates..." notification — no more wondering whether the click did anything. The button restores itself when the check finishes.
- Cloud Storage text no longer ghosts (Windows app): the storage status line and note now repaint opaquely, so old text ("Contacting server...") can't smear underneath the new text.
Version 1.38.9
- Cloud Storage page fixes (Windows app): the Upload / Back up now / Download / Delete / Refresh buttons and the "Add storage" button are visible again — owner-drawn buttons whose IDs were never registered with the main window's painter were painting nothing (invisible but still clickable). Any future content button now falls back to the standard themed painter automatically, so this class of bug can't recur.
- Readable storage text (Windows app): the "Add storage (GB):" label no longer touches the GB input box (proper spacing), and the garbled
â€" characters in the storage status text are gone — non-ASCII dashes in Windows UI strings have been replaced with plain hyphens (tray.c is now pure ASCII). - Managed protection card (My Account → Add-ons): the card no longer shows literal
\n / \u{1f6e1} text — it renders the emoji and clean line breaks in both the purchasable and under-maintenance states. - Fixed a control-ID collision (
IDC_BTN_CONNRETRY shared 639 with the storage Delete button; now 643).
Version 1.38.8
- Version strings now name their component: the Windows app reports and displays
1.38.8-win, the server reports 1.38.8-srv (admin status page), so it's always clear which piece a version refers to. Linux (-lin) and Android (-and) tags are reserved for those builds.
Version 1.38.7
- One dark theme for every tray window: all dialogs opened from the tray (AI chat, AI Skills Checker, firewall rules, dark-web check, document audit, track cleaner, email scan, disk cleanup, backup, vault, LAN scan, bank mode) now share the main window's dark theme — near-black background, light text, cyan-accented buttons. New shared
src/ui_theme.h. - AI Skills Checker (Windows app): new dialog in the tray app that verifies each AI Security Assistant skill in order — add-on status, server config, remaining daily quota, live question-answering, and live detection explanation. Open it from the tray menu ("AI Skills Checker...") or the "Check skills" button in the AI chat window. The two live tests use 2 of the 50 daily questions; the status checks are free. New
POST /api/ai/skills-status endpoint (no quota spent, no AI call). - Licenses are now per device type: a Windows (Personal/Plus/Family/Business) key no longer activates Android or Linux installs — each installer has its own license type (new Android and Linux plans, $2.99/mo for 1 device). Trial keys are scoped to the device type they were created on. The admin panel shows each key's device type and can issue cross-platform keys when needed.
- Stripe setup:
STRIPE_PRICE_ID_ANDROID + STRIPE_PRICE_ID_LINUX (recurring $2.99/mo prices you create in Stripe) enable the new Android/Linux plan cards; without them the cards show honestly as unavailable.
Version 1.38.6
- Managed protection (paid add-on): customers can opt in to having us manage their devices for a small monthly fee (
STRIPE_MANAGED_PRICE_ID) — settings are locked on every device and each device is monitored. They keep ALL features and protections either way; self-manage by simply not adding it. Card on My Account → Add-ons, Stripe checkout, webhook activation, admin grant/revoke, and heartbeat-enforced lock_settings. - User guide: a full documentation page at /docs explains how to use every feature — protection, privacy, performance, cloud storage, family controls, network, account, add-ons, and business tools. Linked in the site footer's Support column.
- Managed protection: the homepage now explains the service is fully manageable — settings can be locked on your devices, and we monitor and protect them for you.
Version 1.38.4
- Rename discovered devices: hostnames in Device discovery are now editable — by admins in the admin panel and by customers from My Account → More → "Devices on your network." Custom names are never overwritten by later network scans; clear one to go back to the detected name.
Version 1.38.3
- My licenses is now My Account: the customer account section (page title, nav, footer, and help pages) has been renamed to My Account.
Version 1.38.2
- Tidier account page: My licenses is now organized into tabs — Licenses, Add-ons, Billing, Security, and More — with every paid add-on in its own Add-ons tab. Deep links like /account#addons still work.
Version 1.38.1
- Simpler plans: the mobile web app is no longer bundled with any plan. Plan pages and the feature list now show exactly what each plan includes — nothing more.
Version 1.38.0
- Cloud storage (paid add-on, per-GB/month): private cloud storage for files and cloud backups. New Cloud Storage page in the app with a live usage monitor, file upload/download/delete, and one-click upgrades straight from the UI. Storage can only be increased, never decreased — stated in the app, at checkout, in the FAQ, and in the ToS.
- Disk Cleanup (paid add-on, one-time): deep-clean temp files, browser caches, Recycle Bin, and thumbnail caches from Performance → Disk Cleanup. Shows exactly what it found before cleaning; only safe, regenerable locations. Unlock from the app or your account page.
- Admin panel: new Cloud Storage section — per-account quota/usage, file inspection and deletion, quota override (the only way a quota can go down), pricing configuration status, and Disk Cleanup grant/revoke.
- Stripe: per-GB subscription quantity syncs to the quota via webhook (increase-only enforced); Disk Cleanup is a one-time payment unlock.
- ToS, Help & FAQ, homepage, and admin guide all updated for the new features.
- New ESET-style site footer (Products / Support / Legal / Account columns, every link working) plus a new Privacy Policy page (/privacy, incl. cookie policy).
Version 1.37.11
- Reinstalls and updates no longer create duplicate devices: registration now recognizes the PC by a stable device fingerprint and reuses its existing device row — even if the old entry heartbeated minutes ago. Your single-seat trial key won't get burned by a stale row anymore, so installs register on the first try.
- Entering your license key in Settings now heartbeats immediately — the status light turns green right away instead of sitting yellow.
- Tray mode now heartbeats on the same schedule as service mode (every 15 minutes by default), so the connection stays live without needing to click anything.
- The server now always answers heartbeats with JSON (never an HTML error page), so a server-side hiccup shows a readable message instead of a stuck yellow light.
- Device rename (two-way sync): rename any device from the admin install detail view or right on the PC in Settings → General → Device name. The newest rename wins and syncs the other way on the next heartbeat (clear the name to go back to the computer name).
Version 1.37.10
- Forensic File Sandbox: detonate a quarantined file in an isolated sandbox on the PC (restricted job object at Low integrity, ~30s). Static scan + AI heuristic + ML behavior scoring of every spawned process. Report uploads to the admin panel automatically.
- Forensic URL Sandbox: analyze a suspicious URL without executing anything — redirect chain, response headers, content SHA-256, static scan of the payload, and extracted indicators.
- New admin "Forensic sandbox" section under Security operations: queue file/URL analyses per device and browse full forensic reports with verdict badges.
- Bigger main window (1120x800, was 1000x680) with centered content — the server connection detail label and Retry button are now fully visible; bigger Settings window (1040x740) with centered tabs.
- Self-healing registration: a client that never got credentials (e.g. rejected at install time) now retries registration every 5 minutes instead of staying disconnected forever.
- Fixed a heartbeat crash (undefined
policy variable) that returned HTTP 500 on every heartbeat — caught by the client test suite before ship.
# AEAntivirus 1.37.9
Version 1.37.9
- Server connection diagnostics: the Status page now shows the exact connection state/error under the server status light, with a Retry button to force re-registration + heartbeat.
- Settings window now has a horizontal scrollbar for wide content.
# AEAntivirus 1.37.8
Version 1.37.8
- Background scans: scans continue in the Windows service if the UI is closed; reopen to see live progress or final results.
# AEAntivirus 1.37.6
Version 1.37.6
- HIPS (Host Intrusion Prevention System): real-time monitoring of new processes.
- ML Behavior Analysis: weighted feature model scores runtime behavior 0-100; malicious processes terminated automatically.
- HIPS/ML status tiles in the admin panel.
# AEAntivirus 1.37.5
Version 1.37.5
- Unprotected Device Discovery: clients scan their local network via ARP and report devices; admin panel shows them with Protected/Unprotected status.
# AEAntivirus 1.37.4
Version 1.37.4
- Device isolation and process blocking in the admin panel install detail view.
block-process:<name> and unblock-process:<name> remote commands.- Straightened-up homepage with collapsible feature sections.
# AEAntivirus 1.37.3
Version 1.37.3
- Force heartbeat: admin panel Heartbeat tile now has a "Force heartbeat" button to make a device check in immediately.
- CPU and memory stats: client reports CPU % and memory MB in every heartbeat; shown as tiles in the admin panel install detail view.
- Low resource usage: client runs at below-normal priority, scan threads deprioritized.
- Responsive web console: admin panel now fully mobile-friendly with bottom nav on phones.
- "Low CPU and memory usage" messaging on website, installer, and main UI status page.
# AEAntivirus 1.37.0
Version 1.37.0
- Completely redesigned Settings UI: clean 6-tab interface (General, Protection, Privacy, Network, Schedule, Lists).
- Every setting now properly loads its saved value and persists on Save.
- Quarantine restore data now survives uninstall/reinstall via registry backup.
- "Made in USA" in app sidebar.
# AEAntivirus 1.36.6
Version 1.36.6
- Scan progress ring: the main circle now fills smoothly as the scan progresses, showing percentage and threats stopped.
# AEAntivirus 1.36.5
Version 1.36.5
- Automatic update checks every 24 hours (both service and UI).
# AEAntivirus 1.36.4
Version 1.36.4
- Fixed tile ID conflicts (Dark Web Scanner, Audit Documents, Clean Browsing Tracks now have unique IDs).
- Fixed scan buttons being cut off (narrowed to fit).
- Removed all UTF-8 garbage characters from notifications and UI (replaced with ASCII).
# AEAntivirus 1.36.2
Version 1.36.2
- Fixed Privacy tile names: Dark Web Scanner and Audit Documents now display correctly (was showing status text due to ID conflict).
# AEAntivirus 1.36.1
Version 1.36.1
- Fixed blank Privacy tiles: Dark Web Scanner, Audit Documents, and Clean Browsing Tracks now display correctly.
# AEAntivirus 1.36.0
Version 1.36.0
- Real-Time + Cloud-Enabled Scans: new "Cloud-enabled scans" toggle in Protection settings. When on, real-time and on-demand scans check unknown file hashes against cloud threat-intel before local heuristics; detections from the cloud carry a "Cloud" badge. Turn it off for local-only scanning.
- Customizable Firewall: user-defined rules (allow/block by application, port, IP/CIDR) in Protection settings, checked before the default firewall logic. First match wins.
- Dark Web Scanner: check if your email appears in known data breaches from the Privacy tab. Only a one-way hash of the address is ever sent — never the address itself — and only with your explicit consent.
- Sensitive Document Auditing: scan your Documents and Desktop for SSN, credit card, and bank account patterns from the Privacy tab. Shows file paths and pattern types only — never the numbers — with secure shredding for flagged files.
- Browsing Track Cleaner: clear Chrome, Edge, and Firefox history, cookies, and cache from the Privacy tab. Prompts you to close running browsers first; never touches passwords.
- Email Anti-Spam & Phishing Protection: new "Email phishing protection" toggle blocks phishing sites via an extended blocklist, plus a "Scan Email for Phishing" tool that checks Outlook and Thunderbird mail for spoofed senders, mismatched links, and urgent lures — quarantining copies without touching your originals.
# AEAntivirus 1.35.68
Version 1.35.68
- Fixed Settings tabs: each tab now shows only its own section (General, Scans, Protection, Privacy, Advanced). Previously all sections were visible at once due to the tab filter running before the window existed and incorrect Y-bounds.
- Settings tabs get the luxury glossy treatment with cyan underline on the active tab.
- Save/Cancel buttons now stay visible on every tab.
# AEAntivirus 1.35.67
Version 1.35.67
- Luxury UI makeover: glossy gradient tile cards with drop shadows and top highlights, glassy status orb with specular shine and halo glow, smooth rounded gradient buttons with cyan hover glow.
- "Server status" label now centered over the status light.
- Settings Save no longer restarts the UI — the dialog just closes with a "Settings saved" note (1.35.66, folded into this build).
# AEAntivirus 1.35.65
Version 1.35.65
- Main screen: the green check now becomes a scan progress indicator while a scan runs - a spinning cyan ring with the live file count in the center, switching back to the checkmark when the scan finishes.
# AEAntivirus 1.35.64
Version 1.35.64
- Settings dialog now uses individual tabs (General, Scans, Protection, Privacy, Advanced) across the top instead of the sidebar.
Version 1.35.62
- Detection pipeline: refined heuristics with off/low/medium/high levels, Authenticode digital signature checks (trusted signatures skip heuristics), monitor-only mode (logs without quarantining).
Version 1.35.60
- Quarantine restore auto-adds an exception: restored files are whitelisted by SHA-256 so they are never detected again.
- Quarantine restore auto-adds an exception: restored files are whitelisted by SHA-256 so they are never detected again.
Version 1.35.59
- Remote input via HTTP: admin clicks, drags, and keyboard work even when the viewer WebSocket is down.
- Remote input via HTTP: admin clicks and keyboard work even when the viewer WebSocket is down.
Version 1.35.57
- Remote file access: admin can list directories and download files from the remote device via the Files button in the viewer.
- Remote file access: admin can list directories and download files from the remote device.
Version 1.35.54
- Server status light: 32x32 gradient orb (green/yellow/red) on the Status page.
# AEAntivirus 1.35.40
Cross-platform antivirus service: Windows (service or system-tray UI), Linux (systemd daemon), Android (app, preview). One subscription key covers every device on the plan, up to its seat limit — seats are enforced across platforms.
Client versions: Windows 1.35.64, Linux 1.21.0, Android 1.21.0 | Server: 1.34.3
- Professional status light: gradient + highlight + bezel to match dark UI.
Version 1.35.53
- Moved server status light down 10px.
Version 1.35.52
- Server status: big light on the right with "Server status" label above, no text.
Version 1.35.51
- Fixed server status light drawing (was hooked to wrong window proc).
Version 1.35.50
- Server status light in the UI: green when connected, yellow while connecting, red on failure.
- Remote support: device sends frames via HTTPS POST instead of WebSocket (WinHTTP WS upgrade fails with 500).
Version 1.35.49
- Remote support: device sends frames via HTTPS POST instead of WebSocket (WinHTTP WS upgrade fails with 500).
Version 1.35.48
- Remote support: use WinHTTP native WebSocket API (WinHttpWebSocketSend/Receive) instead of manual framing.
Version 1.35.47
- Remote support: log the HTTP 500 response body to diagnose the WebSocket upgrade failure.
Version 1.35.46
- Fixed remote support WebSocket: set WINHTTP_OPTION_UPGRADE_TO_WEB_SOCKET so the server returns 101 instead of 500.
Version 1.35.45
- Fixed remote support crash: the WebSocket's hand-rolled Schannel TLS crashed with 0xC0000005 in SspiCli.dll, killing the tray. The transport now uses WinHTTP (same as the rest of the app).
Version 1.35.44
- Remote support diagnostics part 3: worker thread start/exit is now logged, to track down the UI disappearing on Connect.
Version 1.35.43
- Fixed remote support screen capture: the JPEG encoder lookup was failing on some systems. It now uses the standard JPEG encoder CLSID directly.
Version 1.35.42
- Remote support diagnostics part 2: the screen capture now logs the exact step that fails (GetDC, BitBlt, GDI+ bitmap, JPEG encoder, temp file).
Version 1.35.41
- Remote support diagnostics: the pre-flight and service-detection now log exactly why they fail, so a "reporting unavailable" actually tells us what broke.
Version 1.35.40
- Fixed network dying after a crash: if the service died while the DNS filter was active, the system DNS was left at 127.0.0.1 and the PC looked offline. The service now detects and restores stale crash DNS on startup.
Version 1.35.39
- Fixed program disappearing after reboot: a previous uninstall's pending reboot-delete would wipe a fresh reinstall to the same path. The installer now clears stale reboot-deletions.
- Fixed remote support in service mode: the tray no longer steals remote-start/remote-stop commands (it defers to the service, which launches the user-session streamer). Previously the tray grabbed the command, failed its capture pre-flight, and the banner never appeared.
- Fixed installer registration: the key-only installer (1.35.8+) was calling the session-based register which refuses without a login cookie, so installs never activated. It now falls back to key-based registration.
- Fixed crash when clicking Yes/No on the firewall prompt (a leftover WM_QUIT was killing the tray UI).
- Uninstall now removes ALL files (config, logs, quarantine) instead of leaving them behind.
- Fixed crash when remote support couldn't start (the "unavailable" report now uses WinHTTP instead of the hand-rolled TLS that crashed in SspiCli.dll).
- Fixed garbled dash character in the registration error message.
- The server URL fallback is now https://aeantivirus.com instead of empty: a build made without the SERVER_URL env var can no longer silently run in standalone mode (never registering, never appearing in the admin panel).
Version 1.35.34
- Fixed 403 recovery loop (1.35.33 bug): the client now forces a real re-registration when the server rejects its credentials, and the command poll retries it (throttled) so the device reappears in the admin panel within a minute instead of looping forever.
Version 1.35.33
- Service and tray share machine-wide config (fixes credential loss).
- Remote-support banner fix: the client was silently dropping remote-start/remote-stop commands during poll parsing, so the banner never appeared. The parser now accepts both.
- 403 responses no longer wipe credentials before re-registering.
Version 1.35.32
- Command poll logs HTTP failures.
Version 1.35.31
- Server: commands delivered via /api/commands only (no heartbeat steal).
Version 1.35.30
- Server: remote-start now delivered via /api/commands (VALID_CMDS fix).
Version 1.35.29
- Service has dedicated 3s command poll thread (remote support works without UI open).
Version 1.35.28
- Diagnostic logging for command poll.
Version 1.35.27
- DNS filter exclusion is case-insensitive (fixes error 12007).
Version 1.35.26
- Installer kills tray UI processes before copying files (fixes "could not copy program files").
Version 1.35.25
- Reverted background poll thread (stability): uses the 3s UI timer.
Version 1.35.24
- Remote support works reliably in UI mode: dedicated background thread polls for commands.
Version 1.35.23
- Remote support works in service mode: the background service now polls for remote commands.
Version 1.35.22
- Instant tab switches: no slide animation — content stays perfectly centered.
Version 1.35.21
- Refined tab transitions: subtler 16px slide with exact final positioning — content stays centered.
Version 1.35.20
- Fixed page slide transitions (controls no longer jump to wrong positions).
- Admin: clear command history, quarantine records, and detections per device.
Version 1.35.19
- Refined UI transitions: pages slide in smoothly, window opens with a subtle slide-up — no fade effects. Buttons have smooth hover highlighting.
Version 1.35.18
- Fast command polling: the client checks for remote commands every 3 seconds (was 15 minutes via heartbeat), so remote support sessions connect promptly.
Version 1.35.17
- Smooth UI transitions: the main window fades in, pages crossfade when switching, dialogs fade in, and buttons smoothly highlight on hover.
- Unlimited remote support sessions: specific PCs (like DESKTOP-168LQHE) can be marked unlimited in the admin panel — sessions never bill.
Version 1.35.16
- Fixed quarantine delete: items whose files are already gone from disk now clear from the list instead of failing with "file is missing" forever.
Version 1.35.15
- Fixed license key not taking effect: entering the key in Settings now updates the app immediately and re-registers right away (previously the key was saved to disk but the running app kept using the old value until restart).
Version 1.35.14
- Settings unlocked: when the client isn't registered, the policy lock no longer blocks Settings changes.
Version 1.35.13
- Clearer errors: the Status page now shows the server's exact error when registration fails.
Version 1.35.12
- DNS self-protection: the DNS filter can never block aeantivirus.com.
Version 1.35.11
- Self-healing registration: if the server rejects stale credentials, the client automatically re-registers.
Version 1.35.10
- Better diagnostics: the Status page now shows the HTTP status when the server rejects a request.
Version 1.35.9
- Version display: the current version now shows in the window title bar.
Version 1.35.8
- Simpler installer: no sign-in needed — your license key is already linked to your account from the download page. Just enter the key.
Version 1.35.7
- Connection diagnostics: TLS 1.3 support and the actual Windows error code when the server cannot be reached.
Version 1.35.6
- Settings easier to read: Settings text is now larger (11pt body, 12pt bold section headers) for better readability.
Version 1.35.5
- Server connection status: the Status page now shows whether the client is registered, the server URL, last contact time, and any connection error — so registration issues are visible instead of silent.
Version 1.35.4
- No more duplicate tray icons: the installer now opens the existing tray window instead of launching a second copy (service mode keeps its single notifier icon).
- Admin panel Avast-style sidebar: the web admin now uses the same left-sidebar layout as the Windows app.
Version 1.35.3
- Server registration fixed: installing in tray mode no longer loses the server registration when the installer switches data folders — existing broken clients self-repair on next start and appear in the admin panel.
- Quarantine Delete fixed: now finds files in both tray/service folders, asks for confirmation, and reports results.
- Check for Updates button is back on the Status page (it was dropped in the 1.35.0 redesign).
- Settings lock explains itself: when fleet policy locks Settings, the dialog now tells you to open Admin → Fleet → Enforcement and uncheck "Lock Settings".
Version 1.35.2
- Quarantine Restore fixed: restore now looks for the quarantined file in both the tray and service data folders (they use different locations), and shows the actual reason if a restore fails instead of guessing.
- Avast-style installer: the Setup window now has the same dark sidebar with AEAntivirus branding as the main app.
- Auto-run after install: the program opens automatically when installation finishes.
Version 1.35.1
- Settings sync both ways, immediately: toggling Real-time, Firewall, Web Protection, or Advanced RTP now pushes the change to the server right away (single-flight, no stacked heartbeats) instead of waiting up to 15 minutes. "Check for updates" also pulls fresh server policy, and the UI refreshes automatically when policy changes module states.
- Quarantine Restore feedback: the Restore button now tells you if nothing was selected, and reports how many items restored vs failed.
- Avast-style Settings: the Settings dialog now has the same dark sidebar navigation (General, Scans, Protection, Privacy, Advanced) — click a section to jump straight to it.
Version 1.35.0
- Avast-style UI redesign: The main window now has a dark sidebar with navigation (Status, Protection, Privacy, Performance, Settings), a big status circle with checkmark on the Status page, and clickable tile cards for quick actions.
- Status page shows protection state at a glance with tiles for Smart Scan, Core Protection, Quarantine, and Network Monitor.
- Protection page groups all protection toggles and scan controls. Privacy page has Web Protection, Firewall, Encrypted Vault, and Data Shredder. Performance page shows statistics and recent scans.
Version 1.34.6
Files Scanned resets on window open + Settings text blur fix
- 🔄 Files Scanned zeroes every time the main window opens — closing the window only hides it (the app keeps running in the tray), so the counter now resets whenever the window is shown, not just on a full restart. The live count is preserved if a scan is running.
- 🔧 Settings text blur fixed — Settings labels painted with a transparent background, which smeared/ghosted on repaint and scroll; they now paint opaque for crisp text.
Version 1.34.5
Files Scanned resets on restart + Recent Scans blur fix
- 🔄 Files Scanned zeroes on restart — the Files Scanned counter is now a per-session number: it always starts at 0 when the app launches instead of carrying the previous session's count forward.
- 🔧 Recent Scans blur fixed — the 8 Recent Scans history rows smeared when their text updated; they now paint opaque like the live scan line so each update erases cleanly.
Version 1.34.4
Windows UI readability pass
- 🔍 Larger fonts everywhere — body text 9pt → 10pt, bold headers 11pt → 12pt, section headers 9pt → 11pt bold, stat values 14pt → 16pt bold.
- 📐 Bigger buttons, more spacing — action buttons grew from 32px to 36px tall with more vertical spacing between rows; the main and Settings windows are larger.
- 🎯 Centered stat cards — Signatures Loaded, Files Scanned, Threats Stopped, and Lifetime Threats values and captions are centered in their own boxes.
- ✨ Clearer text rendering — ClearType replaced with grayscale antialiasing for the dark theme, removing the fuzzy/blurry look; swept for overlapping text.
- ⏯️ Pause / Cancel Scan buttons — Pause toggles to Resume; Cancel stops the scan; both enabled only while a scan is running.
- 🧹 Recent Scans clear on open — the scan history is wiped every time the main window opens.
- 🔄 Updater opens the main window — after an in-app update restart, the main program window appears automatically.
Version 1.34.3
- Free-trial key fix: since the installer asks for a license key up front, fresh free accounts now get their 7-day trial key created automatically the first time they visit the download or account page — and the key is shown right on the download page. No more "no licenses yet" dead end.
Version 1.34.2
Terms of Service + License Agreement acceptance in all installers
- 📜 ToS and License Agreement with Accept/Decline — every installer now shows the Terms of Service and the End User License Agreement before installing. You must accept both to proceed; declining aborts the install.
- 💻 Windows installer — two checkboxes ("I accept the Terms of Service" / "I accept the License Agreement") with View buttons that open the full document in a dialog with Accept/Decline. The Install button stays disabled until both are checked.
- 🐧 Linux installer — fetches and displays both documents from the server, prompts for acceptance. Set
AEA_ACCEPT_TOS=1 and AEA_ACCEPT_EULA=1 for non-interactive fleet installs. - 🍎 macOS installer — same acceptance flow as Linux.
- 📱 Android app — onboarding shows both documents with checkboxes; the Activate button is blocked until both are accepted.
- 🌐 New endpoints —
/api/terms-text and /api/license-text serve plain-text versions for installers; /license serves the human-readable License Agreement page.
Version 1.34.1
Windows UI polish + in-app updater fix
- 🔧 Blurred scan text fixed — the live "Scanning:" line smeared because it repainted with a transparent background; it now paints opaque so rapid updates erase cleanly.
- 🧹 Garbage characters removed — swept every Windows UI string for non-ASCII (em dashes, middle dots, smart quotes, ellipsis, bullets) and replaced with plain ASCII.
- 📋 Quarantine upgrades — horizontal scrollbar for long paths, Select All button, multi-select with Ctrl+click/Shift+click; Restore/Delete act on all selected.
- 🔄 Updater version fix — the manifest version is now bumped with each build so "Check for updates" actually detects and downloads new builds.
Version 1.34.0
Per-install platform + mode tracking, heartbeat data fix
- 💻 UI / Service / App mode badges — the admin panel now shows *how* each device is installed: UI (Windows tray app), Service (Windows service or Linux daemon), or App (Android). A new Mode column appears in the fleet installs table and per-customer install lists, so you can see at a glance which PCs run the tray UI vs the background service.
- 📱 Platform + mode on every install — Windows, Linux, and Android clients now report both their platform and install mode at registration and on every heartbeat. Older clients show a dash until they update.
- 🔧 Heartbeat data corruption fixed — a SQL argument-order bug was writing the protection-paused flag into the IP column, the pause timestamp into the paused flag, and the IP address into the pause-expiry column. All three now land in the right columns, so the install map and paused-protection indicators show correct data.
Version 1.33.0
Owner-only admin, trial auto-conversion, license key fixes
- 🔒 Admin panel is owner-only — the read-only viewer role (
VIEWER_PASSWORD) has been removed. Only the owner password can sign in to the admin panel; viewer passwords are rejected. - 🔄 Trials auto-delete on paid upgrade — when a customer buys a paid plan, their trial license is automatically deleted and their PCs are moved onto the new paid key. No more orphaned trials, and the new key immediately shows the customer's PCs.
- 💻 "No PCs using this key" fixed — the trial-to-paid migration means PCs registered under a trial key now appear under the paid key after purchase.
Version 1.32.0
Remote screen viewing & control streams (server 1.32.0; Windows 1.29.0, Linux 1.20.0, Android 1.20.0)
- 🖥️ Live screen streaming, all three platforms — starting a remote session now wakes the device over the command queue (
remote-start:<session_id>) and the device opens a TLS WebSocket to /api/remote/stream, streaming JPEG frames (≤512 KiB, ≤10 fps) that the server relays to the owner's browser viewer at /admin/remote-view. No frames are ever stored — the relay only forwards. - Windows 1.29.0 (unattended): GDI BitBlt capture, GDI+ JPEG, SendInput keyboard/mouse, persistent topmost "A support agent is connected" banner with Disconnect button.
- Linux 1.20.0 (unattended, X11 only): X11/XShm capture, libjpeg encoding, XTest input, persistent on-screen indicator with Disconnect. Wayland sessions honestly refuse — the daemon reports
wayland and the session ends unbilled. - Android 1.20.0 (attended-only): MediaProjection consent is required for every session — deny it and nothing connects. Input needs the accessibility service the customer enables manually; without it the session is view-only. Foreground-service notification: "A support agent is connected — tap to disconnect".
- 🔍 Live scan progress — what file, how long left — scans now show the file being scanned as it happens, plus an estimated time remaining. Windows does a fast pre-count pass ("Counting files…") so the ETA is real; Android shows
done / total — filename — ~2m 15s left in the app. - 💳 Never billed for a session that can't stream — new
POST /api/remote/unavailable lets the device report it can't honor a remote-start (Wayland, headless, denied consent); the session ends as no_device, excluded from the 3-free/month quota, and any captured $75 charge is refunded automatically. A 3-minute connect-grace sweep ends silent sessions the same way, and a waiting viewer is told exactly why. - 🔒 Idle sign-out — customer, admin, and reseller browser sessions now expire after 30 minutes of inactivity (activity keeps them alive; install-token/device auth is unaffected).
- 🚨 Streaming safeguards — one active session per PC, one device + one viewer socket per PC, 60 s no-frame timeout, per-second frame/input rate caps, schema-validated input, owner emergency disable (
/api/admin/remote/disable) that ends live sessions and blocks new ones. - Covered by tests —
server/tests/server/v1320-test.js: real HTTP/WebSocket round trips for idle logout, hello/info, byte-identical frame relay, input validation, duplicate refusal, one-session-per-PC, remote-start/stop queueing, emergency disable, no-device refunds, connect-grace sweep, and the add-on card escape-fix regression.
Version 1.31.1
Remote support wording (server 1.31.1; Windows stays 1.27.0, Android 1.18.0, Linux 1.19.0)
- 🙋 Generic "support agent" wording — customer-facing remote-support copy (homepage, account page, Help) now says "a support agent" instead of naming Austin. Feature set unchanged from 1.31.0.
Version 1.31.0
Remote support sessions (server 1.31.0; Windows stays 1.27.0, Android 1.18.0, Linux 1.19.0)
- 🖥️ Pre-authorized unattended remote access — authorize a customer's PC at their request from the new Fleet › Remote support admin section; the customer can revoke anytime from their account page. Every session shows a persistent on-screen "connected" indicator on the PC (no silent sessions), and every session is audit-logged and visible to both sides.
- 💳 Per-session billing: 3 free/month, then $75 — the first 3 sessions per account each calendar month are free; further sessions are $75 each, charged when the session starts. A card on file (saved via the account page) is required for billable sessions; free sessions always work. The owner can waive any session (goodwill), and waived sessions don't consume the free quota.
- 🙋 Customer self-service — the account page shows remote-access status per PC, this month's usage, card-on-file status, and full session history.
- Honest rollout — the Connect button stays disabled under an "under maintenance" banner until screen/input streaming ships in the next release; authorization, session logging, and billing are live now. Terms and Help disclose the capability and pricing.
Version 1.30.0 (admin-only)
Monthly stale-file cleanup (server 1.30.0; Windows stays 1.27.0, Android 1.18.0, Linux 1.19.0)
- 🧹 Monthly stale-file cleanup — new card in System › Maintenance (off by default, same honesty rule as Data retention): when enabled it runs once per calendar month and removes old backups beyond the keep count, crash leftovers from interrupted restores, and stale
*.tmp/*.part scratch files. It never touches the live database, quarantine, branding, installers, or your newest backup. - Preview before anything is deleted — the card shows exactly which files would go and why, plus a one-click Clean now; every scheduled and manual run is audit-logged with counts and bytes freed.
- The daily retention purge (DB rows), the 30-day quarantine file pruning, and the automatic-backup keep count are unchanged — the monthly job covers the gaps none of them cleaned.
Version 1.29.0
Full reorganization: admin panel, homepage, nav, and repo layout (server 1.29.0; Windows stays 1.27.0, Android 1.18.0, Linux 1.19.0)
- Canonical feature names — one name everywhere (admin, homepage, Help, Terms): AI Security Assistant, Identity Theft Monitoring, VPN (WireGuard), XDR incidents, EDR, Centralized console. The admin "Alerts" tab is now Notifications; "System › Security" is now Admin security (tab keys unchanged, so saved URLs keep working).
- Role-aware header nav — customers see My licenses · Add-ons · Download · Help; logged-out visitors no longer see Admin/Reseller links; the reseller portal keeps its own nav.
- Admin "🧩 Add-ons" tab — VPN (WireGuard), Identity Theft Monitoring, and AI Security Assistant each get their own section instead of living under Customers.
- Admin "🚨 Security operations" tab — EDR, XDR incidents, network monitor, detections, suspicious files, quarantine, and SOC/SIEM in one place; Customers/Fleet trimmed to customer- and fleet-management only.
- Section merges — "Threat intelligence" folded into Protection › Feed sources; "Software inventory" folded into Patch management; the dead
vulnerable-software API alias and unused vuln helpers removed. - Policy section grouped — 56 policy controls organized into 5 collapsible groups (Protection / Network / Device control / Updates & patching / Privacy) instead of one long list.
- Help & FAQ — new "Paid add-ons & advanced features" group answering the AI assistant, add-on signup/cancel, Identity Theft Monitoring, VPN, Centralized console, and EDR questions.
- 3-step checklists —
/success and /addon-success now show a 3-step "you're all set" checklist with per-add-on next steps. - Homepage reorder + anchor TOC — "Everything included" now flows Protection → Privacy → Network → Family → Fleet & IT admin → Paid add-ons → Account → System tools → Resellers, with a jump-to anchor table of contents.
- Repo layout — test files moved to
server/tests/ (server/, client/, fixtures/); the Linux tarball moved to dist/ with a root .gitignore; first increment of the server.js split: public page views extracted to server/views/*.js (esc → server/views/util.js), route handlers build a ctx and call the view — every extracted page verified byte-identical against the pre-split render. - Covered by tests — new
server/tests/server/v1290-test.js (canonical names on homepage/help/terms, new admin tabs render, merged sections, grouped policy, FAQ group, checklists, homepage TOC, view-module round trips); all older suites re-run green from their new locations, inline-script regression green.
Version 1.28.0
AI Security Assistant paid add-on (server 1.28.0; Windows client 1.27.0)
- AI Security Assistant — new
aiassist paid add-on (Stripe price STRIPE_AIASSIST_PRICE_ID), following the existing paid_addons pattern: checkout, webhook activation, account-page card (active / purchasable / maintenance), admin visibility. - Assistant endpoints —
POST /api/ai/chat ({message} → {reply}) and POST /api/ai/explain ({detection_id} → {explanation}); install-auth + add-on gate (402 addon_required), 503 when OPENAI_API_KEY is unset, 429 past 50 queries/day/account (ai_usage table), 502 on provider failure. 30s timeout, never hangs. - Privacy by design — server-side OpenAI chat-completions only (default base
https://api.openai.com/v1, default model gpt-4o-mini, both overridable via AI_API_BASE/AI_MODEL env or admin UI → data/ai.json). Only the user's question — or a detection summary the user explicitly asked to explain — is ever sent; file contents never leave the PC. The key is never logged or echoed. Terms disclose third-party processing and the "AI-assisted guidance, not professional advice" scope. - System prompt — frames the assistant as the AEAntivirus Security Assistant (security Q&A, detection explanations, next steps) and refuses malware-creation / AV-evasion requests.
- Admin "AI assistant" section — subscriber count, today's/total usage, per-account table, provider base/model override UI, one-click test query.
- Covered by tests —
server/v1280-test.js (checkout gating, 402/503/429 paths, stubbed-provider forwarding, rate-limit trip, webhook activation, admin APIs, homepage/terms copy); v1270→v1210 suites + inline-script regression still green.
Version 1.27.0
Centralized management console, EDR, automated isolation, device control, cloud reputation, patch management, SOC/SIEM, XDR incidents, folder encryption (server 1.27.0; Windows client contract frozen for 1.26.0)
- Centralized management console — new
device_groups table with per-group policy overrides (resolution order: per-install > per-account > per-group > global fleet policy). Admin "Device groups" section: create/rename/delete, assign installs, edit overrides as JSON (client keys sanitized via sanitizePolicy; server-only auto_isolate_on_ransomware preserved separately). Remote actions per install or per group — scan, isolate, release — queued into the new device_commands table; clients poll GET /api/commands (cap 20, oldest first) and acknowledge via POST /api/commands/ack. Console shows per-install isolation state + command/ack history. - Automated isolation — server-only
auto_isolate_on_ransomware toggle (global fleet policy + per-group override). When a ransomware detection lands and the flag is on, the server queues an isolate command and logs an EDR isolation event. Terms disclose that isolation cuts network except the server lifeline. - EDR —
POST /api/edr/events intake (≤100/batch, latest 500/install kept). Admin "EDR" section: per-install unified timeline (detections + EDR events + remote commands) and a server-side threat-hunting search (hash/summary text, kind, severity, install, time range; LIMIT 500). - Cloud reputation —
GET /api/reputation?sha256= (strict 64-hex validation, 400 JSON on bad input): malicious if in file_signatures_new, clean if whitelisted, suspicious if in suspicious_files, else unknown. Hashes only — never file contents. - Patch management — the old JSON min-safe-version feed is migrated into the
vulnerable_software(name, min_safe_version) table on boot. Admin "Patch management" section: minimums CRUD + per-device outdated apps from reported software inventory. Client patching uses winget, never auto-reboots. - SOC / SIEM integration — admin "SOC / SIEM" section: syslog target config (host/port/TCP-UDP/severities, stored in
data/syslog.json — no new env vars), CEF-style line format documented in the panel (honest: plain TCP/UDP; TLS needs stunnel), detection-webhook event-type filter (webhook_events server-only policy key), and a new alerts API-token scope serving GET /api/alerts (recent detections + EDR events as JSON). - XDR unified incidents —
incidents table + admin "Incidents" section: correlates endpoint detections + high/critical EDR events per install in a 24h window (open/acknowledged/resolved), with one-click isolate/scan inline. Copy is honest: this correlates AEAntivirus's own telemetry, not third-party email/cloud products. - New policy keys (heartbeat
policy_*): usb_storage (allow|block), auto_isolate, cloud_lookup, patch_auto, encrypt_folders (bounded path list, no drive roots). Defaults match client local defaults. - Multi-layered protection (#10) — copy-only: homepage "Everything included" gains a defense-in-depth framing card; no new feature code.
- Covered by tests —
server/v1270-test.js (commands queue/ack, reputation verdicts + 400s, EDR caps, groups CRUD + overrides, isolate/release flow, auto-isolate trigger, incident correlation, syslog config validation, alert API scopes, homepage/terms copy); v1260→v1210 suites + inline-script regression still green.
Version 1.26.0
Web & phishing protection, Secure DNS, rootkit scans, ransomware rollback (server; Windows client 1.25.0) + two paid add-ons: Identity Theft Monitoring and VPN
- Web & phishing protection (Windows 1.25.0, server) — the Windows DNS filter blocks malicious/phishing domains from a server-managed blocklist (
webfilter_domains), synced to clients every 6 hours via GET /api/webfilter/domains. Matching happens locally on the PC — browsing history never leaves the device. Admin panel: new "Web filter" section under Protection for adding/removing domains (seeded from threat-intel domains on first boot). - Secure DNS (Windows 1.25.0) — DNS-over-HTTPS encryption for lookups, in the base product.
- Rootkit scans (Windows 1.25.0) — deep scans hunting hidden rootkits.
- Ransomware rollback (Windows 1.25.0) — versioned copies of protected folders so ransomware-encrypted files can be restored instead of ransomed.
- Identity Theft Monitoring (paid add-on) — if the account email appears in a newly seen public data breach, the account gets an email with a step-by-step identity-protection checklist (credit freeze, fraud alert, FTC report, password changes). Each account × breach is logged to
identity_alerts and alerted once. Honest scope: breach-driven guidance — not credit-bureau monitoring. Stripe checkout at POST /api/account/addons/identity/checkout (env STRIPE_IDENTITY_PRICE_ID + HIBP_API_KEY); admin "Identity alerts" view under Customers. - VPN (paid add-on) — WireGuard tunnel, one click in the Windows app. Clients call
GET /api/vpn/status and POST /api/vpn/config (402 addon_required without the add-on); the server picks the least-loaded enabled endpoint, assigns a client IP from 10.8.0.0/16, generates an X25519 keypair (Node native crypto — no dependency), revokes the install's old peers, and returns a full .conf (private key returned once, never stored). Admin "VPN" section under Customers: endpoint CRUD (name/host/port/capacity/server public key), peer list with revoke. linux/wg-endpoint-setup.sh provisions an Ubuntu VPS as an endpoint. Live deployment + bandwidth are your step; the panel only stores endpoint addresses. Fair use + minimal-metadata logging policy in the Terms. /addon-success is now addon-aware (?addon=identity|vpn|vault|leakalerts) with per-addon copy.- Covered by tests —
server/v1260-test.js (webfilter sync + admin CRUD, VPN 402 gating + config issuance + key rotation + admin CRUD + revoke, identity checkout gating, addon-status API, admin sections render, homepage/terms copy); inline-script regression and the 1.25.0/1.24.0/1.23.0 suites still green.
Version 1.24.0
Every install needs an account — even a free one (server; Windows/Linux/Android clients updated to match)
- Account required to install —
POST /api/register now returns 401 account_required without a signed-in session, and /download plus the direct file routes (/download/file/:name, /download/AEAntivirus-Setup.exe) redirect anonymous visitors to the free-account card / login. No more anonymous trials: every trial is minted for — and linked to — a real account (email + account_id on the license), and installing with a key attributes that key to the installing account. Existing installs are unaffected (heartbeats and updates don't go through registration). - Free account keeps working — the whole flow costs nothing: create the free account on the site, download, and the installer signs you in (or creates the account for you) before activation. Trial, devices, and license keys all live under the account.
- Clients — Windows installer (1.23.0), Linux
install.sh + daemon (1.18.0), and Android onboarding (1.18.0) now collect the account sign-in during setup and replay the session cookie to the server; passwords are never written to config or logs. - Terms — the free-trial section now states that an account (free is fine) is required for every installation.
- Covered by tests —
server/v1240-test.js (anonymous download/register denials, free-account → trial attribution, key attribution, heartbeat unaffected).
Version 1.23.0
MISP source editing, changelog & status pages, Linux parity (Linux daemon 1.17.0; Windows client 1.22.0 and Android 1.17.0 unchanged; server)
- Edit MISP/STIX sources in the admin panel — each source row has an Edit button that loads the source into the form; saving hits
POST /api/admin/misp/:id/update. Blank auth value keeps the stored credential; a "clear saved credential" checkbox wipes it. Validation and SSRF protection match source creation. Fixes the stuck-bad-URL problem (e.g. a 404ing feed) without delete-and-re-add. - Changelog page — new
/changelog route renders the README's "## New in X.Y.Z" sections newest-first with a "latest" badge; linked from the site footer. No JavaScript. - Public status page — new
/status route shows server/client/daemon/app versions, definition and blocklist versions + entry counts, threat-intel feed health, and MISP/STIX aggregate health. Only public-safe aggregates: no source names, feed URLs, or credentials ever leave the server. Cached 60 seconds; linked from the site footer. - Linux daemon 1.17.0 — MD5/SHA-1 definition parity with Windows (parses v4
md5:/sha1: rows, computes all three hashes, EICAR seeded under all three) plus blocked code-signing certificate parity (Authenticode signer serial/thumbprint matched against certs.db, fail-open). - Homepage — "New in 1.23.0" section.
- Covered by tests —
server/v1230-test.js (changelog/status pages, MISP edit round-trip, Linux source markers).
Version 1.22.0
Help & FAQ (Windows client 1.22.0; Linux 1.16.0 and Android 1.17.0 unchanged; server)
- Help & FAQ on the site — new
/help page with accordion FAQs (Getting started, How protection works, Account & billing, Troubleshooting) plus an administrator note pointing at the MalwareBazaar/YARA/certificate review workflows. Linked from the site header nav (logged-in and logged-out) and the footer. No JavaScript on the page — plain <details> elements. - Help & FAQ in the Windows app — new "Help & FAQ..." item in the tray menu opens a built-in Help window with a read-only FAQ covering scans, quarantine, "suspicious" vs confirmed detections, pausing, updates, false positives, and license keys, plus an "Open online help" button that launches the site's
/help page. - Homepage — "New in 1.22.0" section; new System tools feature card for built-in Help & FAQ.
- Covered by tests —
server/v1220-test.js (/help returns 200 with FAQ markers, nav/footer links present on / and /help, admin link wiring).
Version 1.21.0
MD5/SHA-1 definitions, code-signing certificate blocklist, YARA rules (Windows client 1.21.0; Linux 1.16.0 and Android 1.17.0 unchanged)
- MD5 and SHA-1 virus definitions — definitions now live under
(hash_type, hash) alongside SHA-256. The admin panel's hash-type selector adds MD5 (32 hex) and SHA-1 (40 hex) with per-type validation; signatures.db v4 publishes md5:/sha1:-prefixed rows that older clients safely ignore. The Windows scanner computes all three digests in one pass and checks them in order: SHA-256 → MD5 → SHA-1 → blocked signer certificate → fuzzy. - Code-signing certificate blocklist — admin-managed list of Authenticode signer certificates known to sign malware, published to
server/updates/certs.db (format: <serial-hex> <thumbprint-sha256|-> <subject...>) and mirrored in every platform manifest (certs_url / certs_version / certs_sha256). The Windows client extracts each executable-like file's signer certificate via WinVerifyTrust and matches by normalized serial or SHA-256 DER thumbprint; fail-open by design (unsigned files, missing certs.db, and trust-API failures never break a scan). - One-click MalwareBazaar certificate import — the blocklist card pulls MalwareBazaar's public CSCB export into a review queue (no
MB_AUTH_KEY needed); each staged certificate must be explicitly approved before it goes live. Set CSCB_EXPORT_URL to override the feed in tests. - YARA rule library (server/admin-side) — write, enable/disable, and delete YARA rules from the admin panel, and scan uploaded samples (10 MB cap, never executed or written to disk) against the enabled rules right in the browser. The supported subset (text/hex strings,
nocase/fullword/wide, any/all/N of them) is enforced at save time; unsupported syntax is rejected, never silently misread. Not embedded in the Windows client for this release. - MalwareBazaar hash imports keep every hash — staged imports now parse and retain MD5 and SHA-1 alongside SHA-256, and approving a staged row publishes all three as live definitions.
- Migration — pre-1.21 databases are migrated automatically (old
sha256-PK table → (hash_type, hash); EICAR seeds under all three hash types; old signatures.db rows stay valid). - Covered by tests —
server/v1210-test.js (75 checks: MB MD5/SHA-1 parsing, CSCB parsing, YARA subset parser/scanner, EICAR ×3, definition CRUD/validation/migration/v4 publish, staged approval of all hashes, cert CRUD/staging/publish, YARA CRUD/toggle/sample scan, admin UI markup, old-DB migration); server/certblock-test.sh (30 checks) for the portable C parsing core. - Homepage — "New in 1.21.0" section; new Protection cards for blocked signing certificates and YARA rules; the threat-intelligence card updated.
Version 1.20.0
MalwareBazaar staged import (server-only release — clients unchanged)
- One-click import from MalwareBazaar — Protection → Virus definitions now has an "Import recent from MalwareBazaar" button that pulls the latest 48h feed (hashes only — malware samples are never downloaded).
- Staged for review, never auto-live — only samples with a *named* malware family are staged (unlabeled ones are skipped), up to 500 per import; each staged hash must be explicitly approved before it becomes a live definition and publishes to clients. Rejected hashes never go live. Re-imports skip anything already live or already decided.
- Gated by
MB_AUTH_KEY — with no key set, the import button is disabled behind an under-maintenance banner (same pattern as VAPID push). Get the key free at auth.abuse.ch and set it as a Render environment variable. - Covered by tests —
server/mbimport-test.js (26 checks): CSV parser unit tests (quoted fields, unlabeled/invalid rows skipped, 500-row cap) plus the full endpoint flow against a local fake feed (503 gating, stage → approve → publish, reject, duplicate skipping).
Version 1.19.0
Virus definitions
- Admin-managed file-hash signatures — add SHA-256 definitions from the admin dashboard (Protection → Virus definitions); they're published to
server/updates/signatures.db and pushed to every device on the next update check. Definitions carry severity, category, name, description, and source; all four platform manifests are bumped to the same version with a matching sha256 (the stale macOS manifest heals on first publish). - Honest seed policy — the database seeds with only the genuine harmless EICAR test hash; no fabricated malware hashes are ever shipped. Tested end-to-end by
server/v1190-test.js (add/validate/duplicate/replace/remove, publish integrity, manifest convergence).
AI-assisted detection (Windows client 1.19.0)
- Ten-signal heuristic scorer — the
src/aiscore.c engine now reads writable+executable sections (+15), appended overlay payloads >4KB (+15), TLS callbacks (+10), packer-style section names (+10), and process-injection import patterns (+20 for ≥3 of the VirtualAlloc/WriteProcessMemory/CreateRemoteThread family), on top of entropy, entry-point, section-count, dual-extension, and risky-location checks. - Same safe contract — threshold stays 60, threat name stays
AI.Heuristic.Suspicious, verdicts say *suspicious* (never confirmed malware), and ≥60 still queues metadata to admin triage. - Covered by tests — the pure-C scoring core is marked
AV190-TESTABLE-AISCORE and unit-tested on Linux by server/ai-score-test.c via server/ai-score-test.sh (41 checks: synthetic-PE feature extraction, scoring math, malformed-input safety); compiles warning-free under both MinGW targets.
Homepage — "New in 1.19.0" section; the AI-assisted-detection and threat-intelligence cards updated.
Version 1.18.0
Customer account security
- Passkeys (WebAuthn) — sign in with fingerprint, face, or PIN instead of a password. Phishing-resistant; counts as two-factor on its own. Add/manage from the account page;
@simplewebauthn/server added to dependencies (root + server package.json). - Email verification — new accounts confirm their email (48-char token, 24h expiry) before buying; resend from the account page.
- Active session management — the account page lists every signed-in browser/device (IP, user agent, sign-in time) with per-session revoke and "revoke all others". Only an 8-char token prefix ever leaves the server.
- Breached-password guard — passwords checked against HaveIBeenPwned at signup via k-anonymity (only a 5-char SHA-1 prefix leaves the server; no API key needed). Breached passwords rejected; fails open if HIBP is unreachable.
- Download your data — one click exports licenses, devices, sessions, alert rules, and leak-check records as JSON. Password hashes and secrets never included.
Admin operations
- Encrypted server backups — set a passphrase and backups are encrypted with AES-256-GCM (
AEBK01 envelope: scrypt-derived key, random salt/IV); the plaintext archive is deleted right after encryption so only the encrypted file is kept. Restore/validate accept the passphrase via X-Backup-Passphrase header only. - Backup dry-run validation — new "Validate only" admin action: archive integrity, safe-path check,
aeantivirus.db presence, and full SQLite integrity check in a staging dir — live data untouched, no restart needed. - Audit-log search — admin audit trail filterable by free text, action prefix, and date; CSV export honors the same filters.
Platform hardening
- Global API rate limiting — 600 requests/IP/15min sliding window across all
/api/ routes (429 + Retry-After); existing tighter login/checkout/contact/trial limits unchanged.
Homepage reorganization — the 27-card "Accounts & business" section is now three sections: "Your account & security", "Resellers", and "Fleet & IT admin". "New in 1.18.0" replaces the 1.17.0 highlights; new cards added for passkeys, email verification, session management, breached-password guard, data export, customer 2FA, encrypted backups, dry-run validation, and audit-log search.
Version 1.17.0
Security hardening
- Brute-force protection — customer, admin, and reseller logins lock out for 15 minutes after 5 failed attempts (per IP). Successful logins reset the counter.
- Checkout throttling — max 20 checkout requests per IP per hour, covering promo-code guessing at payment time.
- Reseller 2FA — resellers can enable TOTP authenticator codes with ten one-time scrypt-hashed recovery codes; no session is issued until verification passes. Five bad codes lock the 2FA step for 15 minutes.
- Scoped reseller API tokens — resellers generate
read or commands scoped tokens for their own integrations; plaintext shown once, SHA-256 stored, revocable anytime. Tokens can't touch other resellers' data. - Trial-abuse hardening — Windows and Android clients now send a hashed device fingerprint at registration; trial limits are 3 per fingerprint per 24h (plus the IP backstops). Server re-hashes the received fingerprint — raw device IDs are never stored.
Admin operations
- One-click backup restore — upload a
.tar.gz server backup from the admin panel (owner-only): gzip magic + tar readability checked, path traversal rejected, aeantivirus.db required, SQLite integrity-checked in staging, a pre-restore-safety snapshot taken before the swap. Server restart required after restore. - Web Push for every alert class — push notifications now cover alert-rule firings (critical detections, expiry, offline installs, quarantine retrieval), seat overuse, stale PCs, new/cancelled subscriptions, failed payments, and threat-feed stale/failing. Admins subscribe from the admin Notifications card (shared
/sw.js); controls disable with an under-maintenance banner when VAPID env vars are missing.
Android app 1.17.0
- Remote admin commands — the app now honors the server's command vocabulary from heartbeat:
scan-now starts a scan, update-now checks the manifest, lock-now/unlock-now lock/unlock the app UI with the admin's message (app-level lock — the app can't lock the device), uninstall opens the system uninstall confirmation. reboot-now and quarantine commands are honestly reported as unsupported on Android, never faked. One centralized handler serves all heartbeat paths so commands are never silently dropped. - Build pipeline hardened:
build-manual.sh re-extracts the work-runtime manifest from its AAR if /tmp was wiped; 0 missing library R classes verified in DEX.
Windows client 1.17.0 — x86/x64 installers rebuilt (mingw-w64, Win7-compatible): sends the hashed device fingerprint at trial registration.
Homepage — "New in 1.17.0" section; "Everything included" cards added for brute-force protection, reseller 2FA, scoped API tokens, one-click backup restore, Android remote commands, and full Web Push coverage. Removed the absolute "so you're never unprotected" watchdog wording.
Version 1.16.0
Windows client — tray watchdog
- If the protection tray crashes, a lightweight companion process (
AEAntivirus.exe --watchdog) restarts it automatically — max 5 restarts per 10 minutes, then it backs off and logs. Windows x86/x64 installers rebuilt.
Linux client — parity with Windows
- Scheduled full-system scans (opt-in via
schedscan=1, schedscan_time, schedscan_freq daily/weekly) and the ransomware canary tripwire (decoy files re-hashed every 60s, 3+ changes in 5 minutes trips a critical alert with evidence-gated offender termination). Compiled -Wall -Wextra clean; 38/38 self-tests pass.
Android app 1.16.0
- Launcher icon and welcome screen now use the AE brand logo (matches the website). Notification icon keeps the brand glyph (Android tints it white).
- Includes a crash-report catcher: if the app crashes on launch, reopening it shows the crash details with a Copy button so the cause can be diagnosed from the device.
Server
- Automated backups — scheduled snapshots (configurable hour + retention), one-click manual run, download/delete, full history in the admin panel.
- Promo/coupon codes — admin-created discount codes applied at Stripe checkout (
coupons table, redemption tracking). - Threat-feed health — MISP/STIX sources report healthy/stale/failing with failure counters, alert deduplication, and admin alerts.
- Detection trend charts — admin dashboard charts detections over time.
- Reseller customer health — per-customer healthy/attention/critical classification in the reseller portal.
- Quarantine bulk actions — restore or delete up to 200 quarantined items at once, with per-item results and audit logging. Fixed: nonexistent items are now reported as failures (404), never counted as successes.
- Contact page + Terms of Service —
/contact (5 messages/IP/hour, SMTP or stored) and /terms, linked in the footer.
Homepage — "New in 1.16.0" section; "Everything included" cards added for the tray watchdog, promo codes, detection trends, reseller health, quarantine bulk actions, and scheduled server backups.
Version 1.15.0
Windows client — 2 new protection modules (server-side fleet policy + tray Settings in this release)
- 🗓️ Scheduled full-system scan (opt-in) — pick daily or weekly at a time of your choosing (default 2:00 AM); the engine reuses the existing full-scan path. Never runs while another scan is active (retries on the next scheduler pass). Tray Settings checkbox + fleet policies
policy_schedscan (off by default), policy_schedscan_time ("HH:MM"), policy_schedscan_freq (daily/weekly). - 🪤 Ransomware canary tripwire — hidden decoy files planted in AppData and Documents are re-hashed every 60 seconds; 3 changes inside a 5-minute window trips a critical
Ransomware.Behavior.Canary threat alert with on-screen notification. Optional offender termination uses the Windows Restart Manager to find exactly which processes have a canary file open and terminates only those PIDs (never kills without that evidence; idle/system/self PIDs are never touched). Tray Settings checkbox + fleet policies policy_canary / policy_canary_kill (on by default). - Engine scan mutual-exclusion — a single scan may run at a time: the quick-scan engine function is now guard-claimed, the scheduled days-based scan and scheduled folder scans stamp their last-run date only when they actually run (a skipped scan retries instead of being forgotten), and the tray "Scan now" balloon says "A scan is already running" when a scan is busy instead of falsely reporting completion.
- Fleet policy keys grow from 49 to 54; Windows x86/x64 rebuilt (version resource 1.15.0).
Server — security & billing hardening
- Admin 2FA recovery codes — ten single-use recovery codes (scrypt-hashed, owner-regenerable) as a backup for admin TOTP.
- Subscription alerts — admin email alerts for new subscriptions, cancellations, and failed payments (SMTP-gated).
- Custom-plan billing hardening — per-license serialized adjustments, Stripe-first quantity updates with a fresh idempotency key per attempt, conditional local updates, and atomic registration admission (seat check + grow + install insert under one lock). Covered by 30 mocked-Stripe tests: success, failure-atomicity, concurrent registrations, concurrent +/-, crash/retry, repeated same-target transitions, shrink warnings.
- Trial-signup rate limiting — max 3 trial signups per IP per rolling 24 hours.
- Password Vault v2 (zero-knowledge) — browser-side Web Crypto (PBKDF2-SHA256 600k iterations, AES-GCM); the master password and plaintext entries never leave the device. One-way v1→v2 migration in the vault UI.
- Reseller license CSV export — resellers can download their full key list (key, plan, seats, status, expiry) as a spreadsheet.
- PWA icons — proper 192×192 and 512×512 install icons in the mobile web app manifest.
- Cloud-check verdict language — file-check results say *suspicious* (
AI.Heuristic.Suspicious), never confirmed malware. - Homepage — pricing plans removed from the homepage (they live on
/subscribe); "Everything included" cards added for the mobile web app, cloud file check, and zero-knowledge vault.
macOS client 1.15.0 (source in macos/) — launchd daemon with configurable sweep, on-demand scans, quarantine, and license sync. Source-only: no Mac SDK here, so no compiled/signed binary; needs a real Mac to build and test.
Version 1.14.0
- Mobile web app (PWA) —
/mobile: phone-friendly portal with cloud file check (upload → SHA-256 → threat database), device status, recent alerts. Installable via Add to Home Screen. - Plan gating — mobile web app only on multi-device plans (Plus, Family, Business, Enterprise, Custom, Mobile). Personal plan shows an upgrade prompt.
- New Mobile plan — $5.99/mo — unlimited mobile devices, mobile web app included (
STRIPE_PRICE_ID_MOBILE). - Threat push alerts — Web Push via VAPID (
VAPID_PUBLIC_KEY/VAPID_PRIVATE_KEY); notifies the phone the moment the cloud file check finds a threat. - Password generator — in the Password Vault's Add-entry form: one-tap generation with length and character-set options, built on
crypto.getRandomValues on-device.
Version 1.13.1
- Under-maintenance banners — any button or feature missing its Render environment variable now shows as disabled with an "under maintenance" banner instead of failing silently (per-tier Stripe prices, Password Vault, leaked-credential alerts)
Version 1.13.0
- Plus plan — $7.99/mo for up to 3 PCs
- Enterprise plan — $49.99/mo for up to 25 PCs with priority support
- Custom plan — pick your exact device count at $4.99/device/mo; the charge grows automatically when you install on a new PC and shrinks when you remove one
Version 1.12.0
Password Vault (paid add-on) — encrypted storage for logins and passwords:
- Separate vault master password (distinct from the account password) unlocks the vault at
/vault - AES-256-GCM encryption with a scrypt-derived key; server never stores plaintext passwords
- Add/edit/delete entries (site, username, password, notes); passwords revealed only on demand
- Change-vault-password re-encrypts every entry with the new key
- Stripe subscription add-on (
STRIPE_VAULT_PRICE_ID); 30-minute vault sessions, HttpOnly cookies
Version 1.11.0
Client (Windows) — 6 new capabilities (server-side policy plumbing + remote commands in this release; client binaries ship separately)
- Remote PC lock/unlock — lost or stolen laptop? Queue
lock-now:<message> from the admin panel's install row; the client locks the workstation and shows your message, and unlock-now releases it. Lock state persists across reboots. Fleet policy policy_lockcmd (on by default). - USB device allowlist — only approved USB devices (by hardware serial number) work; everything else is blocked. Fleet policy
policy_usballow (on by default). - Rogue-device network alerts (opt-in) — instant alert when an unrecognized device joins your network. Fleet policy
policy_roguealert (off by default). - Quiet hours (opt-in) — scheduled nightly window (
quiet_start/quiet_end, "HH:MM") during which routine notification popups are suppressed; critical threat alerts always break through. Fleet policy policy_quiethours (off by default). - Notification history — every alert the client showed, kept in a browsable history in the tray UI. Fleet policy
policy_nothistory (on by default). - Protection-paused heartbeat — the client reports
protection_paused / paused_until in every heartbeat so the fleet list shows exactly which PCs are paused.
Server / fleet admin & business
- Quarantine browser — Admin → Fleet → Quarantine lists every quarantined file fleet-wide (filter by filename, threat, hostname, path) with one-click restore (queues
restore_quarantine:<qid>), fetch (downloads the server-side copy), and delete (queues delete_quarantine:<qid>). - Protection-paused fleet alerts — installs that paused protection get a red PAUSED badge in the installs table.
- Monthly customer security digest — on the 1st of every month each customer gets a plain-English email: devices protected, versions, threats blocked in the last 30 days, license status. Toggle via
customer_digest_enabled (on by default); honest no-op when SMTP is unconfigured. - Reseller commission tracking — each reseller gets a commission % (default 20%) and a payout schedule (daily / weekly on Mondays / monthly on the 1st). Commissions accrue on every paid Stripe invoice attributed to the reseller; ready payouts are marked paid manually from Admin → Resellers → Commission payouts (nothing is sent automatically). Resellers see their earnings and payout history in the reseller portal.
Fleet policy keys grow from 42 to 49. All four fleet templates cover the new keys (roguealert and quiethours stay opt-in everywhere).
*Not tested on real Windows yet — installer UI, the new 1.11.0 client features, and the firewall-popup crash still need verification on a physical machine.*
Version 1.10.0
Client (Windows) — 5 new protection modules (server-side policy plumbing in this release; client binaries ship separately)
- Startup-folder persistence monitor — watches the per-user and common Windows Startup folders for new or changed entries (shortcuts resolved to their real targets via the .lnk parser). Registry Run/RunOnce keys were already watched by the registry guard; this closes the remaining classic persistence route. Fleet policy
policy_runmon (on by default). - DNS-over-HTTPS bypass guard — flags or blocks known DoH endpoints so apps can't silently dodge DNS-level filtering through encrypted DNS. Fleet policy
policy_dohguard (on by default). - Hosts-file tamper monitor — watches the system hosts file for modifications, catching malware that redirects banking and email sites. Fleet policy
policy_hostsmon (on by default). - Suspicious-file submission to admin — when the AI heuristic scorer flags a file, the client reports its metadata (SHA-256, filename, size, score — never the file itself) to the server's triage queue. Fleet policy
policy_filesubmit (on by default). - Listening-port auditor (opt-in) — alerts when an unknown program starts accepting network connections, a classic backdoor signal. Off by default (can be noisy). Fleet policy
policy_portmon.
Server / fleet admin & business
- Suspicious-file triage queue — Admin → Suspicious files lists every AI-flagged submission newest-first (SHA-256, filename, AI score, host, account) with a status filter and actions: mark reviewed, dismiss, add note. Submissions are stored in a new
suspicious_files table; client endpoint POST /api/client/suspicious uses the same install auth as the heartbeat. - Stale-install fleet alerts — installs with no heartbeat in 7+ days (
STALE_ALERT_DAYS) get a red STALE badge in the installs table plus a banner count at the top of the section, so silent failures never sit unnoticed. - One-click server backup — Admin → Maintenance → Download backup streams a
.tar.gz snapshot of the server data directory (database, policies, logs); owner-only, audit-logged.
Fleet policy keys grow from 37 to 42. All four fleet templates cover the new modules (portmon stays opt-in everywhere). The admin fleet policy editor now lists every module, including the 1.8.0/1.9.0 ones that were missing.
*Not tested on real Windows yet — installer UI, the new 1.10.0 client features, and the firewall-popup crash still need verification on a physical machine.*
Version 1.9.0
Client (Windows) — 10 new protection modules
- Process-launch behavior guard — watches for suspicious system-tool (LOLBin) misuse and macro-style launches (e.g. Office apps spawning script interpreters). Alerts by default; optional process termination (
procguard_kill, off by default). Settings toggle + fleet policies policy_procguard / policy_procguard_kill. - Scheduled-task persistence monitor — watches for new boot/logon scheduled tasks pointing at suspicious locations, a classic malware persistence trick. Fleet policy
policy_taskmon. - Keylogger activity alerts — always-on heuristic watch for keylogger-shaped programs (no visible window, running from user-writable locations). Alerts only, never terminates. Fleet policy
policy_keylogalert. - ARP spoofing detector — pins your gateway's hardware address and alerts if it changes, catching man-in-the-middle attacks on your local network. Fleet policy
policy_arpguard. - Browser extension auditor — rates Chrome, Edge, Brave, and Firefox extensions by their permissions and flags the ones that can read everything you browse. Fleet policy
policy_extaudit. - Trusted Wi-Fi pinning — learns your home and office networks and warns about evil-twin impostor access points. Fleet policy
policy_wifipin. - WMI persistence monitor — watches the WMI event subscription store for backdoor filters and consumers. Fleet policy
policy_wmimon. - Outbound traffic anomaly monitor (opt-in) — learns each program's normal upload volume and alerts on unusual spikes. Off by default (can be noisy). Fleet policy
policy_exfilmon. - Ad & tracker blocker — DNS-level blocking of ad, tracker, and analytics domains with a built-in conservative list plus your own custom entries. Fleet policy
policy_adblock. - AI-assisted detection — a local weighted-heuristic scorer (NOT machine learning, NOT a cloud verdict) flags files as *suspicious* by structure and entropy when no signature matches. Honest by design: never reports confirmed malware. Fleet policy
policy_aiscore. - One-click malware scan — the tray's "Scan now" runs the existing quick-scan engine, now with the AI-assisted scorer in the pipeline.
Server / fleet admin & business
- Leaked-credential alerts (paid add-on) — weekly checks of the customer account email against the HaveIBeenPwned v3 breach database (requires a paid HIBP API key set as the
HIBP_API_KEY env var on the server; without it the feature reports "not configured" and checks nothing). Customers buy the add-on through Stripe checkout (STRIPE_LEAK_ALERT_PRICE_ID); a dashboard warning naming the breaches and an email alert fire the first time new breaches are found — advisory data only, never invented.
Fleet policy keys grow from 26 to 37. All four fleet templates cover the new modules (exfilmon and procguard_kill stay opt-in everywhere).
*Not tested on real Windows yet — installer UI, the new 1.9.0 client features, and the firewall-popup crash still need verification on a physical machine.*
Version 1.8.0
Client (Windows)
- Ransomware behavior blocker — watches Documents, Desktop, and Pictures for two classic ransomware behaviors the change-burst heuristic misses: mass re-extension renames (10 files renamed to the same new extension within a minute) and mass in-place encryption (15 files turning high-entropy within a minute, 4 KB minimum, common compressed/media formats skipped). Fires a critical alert and stops/quarantines the suspect. Toggleable in Settings (on by default); fleet policy
policy_ransomblock. - Silent / gaming mode — manual tray switch plus auto-engage during fullscreen apps (games, video). Routine popups are dropped outright while silent; firewall and microphone prompts wait politely; critical threat alerts always break through. Fleet policy covers the behavior blocker only — silent mode is the user's own switch.
- Banking keylogger termination (opt-in) — safe banking mode gains an opt-in setting: keylogger-shaped background programs (no visible window, running from a user-writable location) are stopped instead of merely warned about, with a sweep of already-running processes when the session starts. Off by default; every message says it's a behavior guess, not proof. Fleet policy
policy_bankkeylog.
Server / fleet admin & business
- Customer self-service seat release — the
/account page lists each PC by hostname and last check-in with a Remove button, so customers can free their own seats without emailing support. - Trial-expiry emails — optional nurture email before a trial lapses (off by default, SMTP required, 1–7 days configurable, sent once).
- Per-reseller detection webhooks — resellers can each configure their own detection webhook (plus a test button); the global webhook still works as before.
- Reseller revenue report — Admin → Resellers gains a Load report card attributing paid Stripe invoices to resellers (subscription-first, customer fallback), with unattributed direct revenue shown separately.
- Seat-overuse & stale-PC alerts — configurable health checks (on by default, 1–3650-day stale threshold) that raise deduplicated dashboard notifications when a license exceeds its seat count or a PC stops checking in.
Bugfix
- Installing no longer inherits the downloaded Setup's "this file came from the internet" mark — the repeated Windows security warning after double-clicking the desktop icon is gone.
- A crash logger now records exception details to the daily log to help diagnose the reported firewall-popup crash (root cause still under investigation).
*Not tested on real Windows yet — installer UI, the new 1.8.0 client features, and the firewall-popup crash still need verification on a physical machine.*
Client (Windows)
- Encrypted file vault — lock sensitive files in an AES-256-encrypted vault (envelope encryption: a random 256-bit master key, AES-256-CBC with a fresh IV per file; the master key itself is protected by your password via PBKDF2). Auto-locks after inactivity; fleet policy
policy_vault (on by default). - Automatic backup scheduler — pick folders and a daily time; backups run automatically on schedule with snapshot rotation and one-click restore from Settings. Fleet policy
policy_ebackup (off by default). - Third-party software updater — detects outdated third-party applications (14 known vendors in the built-in catalog: Firefox, Thunderbird, Chrome, VLC, 7-Zip, Notepad++, LibreOffice, Python, Git, Java, Adobe Acrobat Reader, WinRAR, FileZilla, PuTTY) and opens the vendor's download page or runs the update for you.
Server / fleet admin & business
- Customer two-factor authentication — customers can enroll TOTP 2FA on their account (authenticator app + 8 recovery codes); login then requires the second factor, with rate-limited verification and lockout on repeated failures.
- White-label branding — Business-tier accounts can rebrand the customer-facing pages with their own product name, logo, and accent color.
- Reseller accounts — create resellers with their own logins, discount percentages, and pricing; resellers get a portal showing their customers and licenses, and can revoke/extend licenses. All reseller actions are audit-logged.
- Data retention controls — configure per-table retention (detections, audit log, heartbeat stats: 1–3650 days or off) with a manual purge and a preview of what would be deleted.
- Install map (approximate) — geographic view of installs from heartbeat IP data; coarse/approximate locations only, with a clear notice that precision is limited.
- Dunning management — configurable failed-payment follow-up schedule (0–90 days) with per-account dunning state tracking; reminder emails go out when SMTP is configured.
- MISP/STIX threat import — connect MISP or STIX/TAXII threat-intel feeds (URL + auth validated, 1–168h pull interval); pulled indicators are merged into the threat feed, with per-source run history.
Bugfix
- Fixed a 1.6.0 admin-page script bug (
x.patterns.join newline SyntaxError) — the admin dashboard JavaScript now parses cleanly.
*Not tested on real Windows yet — installer UI, vault/backup/updater flows, mic/protected-folder/banking prompts, app-limit enforcement, and the LAN scanner still need verification on a physical machine.*
Version 1.6.0
Client protection (Windows)
- Microphone protection — sibling of the webcam guard: detects when a process accesses the microphone, shows an allow/block prompt on first sight per executable, and remembers the choice. Toggleable in Settings; fleet policy
policy_micprotect. - Protected folders — pick folders to protect (defaults: Documents, Pictures); only trusted apps may write there. When an untrusted process writes to a protected folder, the process is stopped, the file is quarantined, and you're alerted with allow/allow-always options. Best-effort by design (no kernel filter driver): it detects and stops untrusted writers rather than blocking the write itself — the Settings page says so plainly. Fleet policy
policy_protectedfolders. - Safe banking mode — tray + Settings toggle. Opens your default browser and shows a "Banking mode ON" indicator while watching for newly started unknown processes and suspicious keyboard hooks. It does not sandbox the browser — the UI copy states exactly what it does and doesn't do. Fleet policy
policy_bankingmode. - Parental app time limits — daily time limit per application (minutes); when the limit is hit the app is closed for the rest of the day with a friendly notice. Usage persists across reboots and resets at midnight. Managed in Settings; fleet policy
policy_applimits (off by default). - LAN device scanner — scans your local subnet with ARP (no port scanning, non-intrusive), lists IP/MAC/hostname, and flags never-seen-before devices as NEW. Manual scan button + optional daily scheduled scan.
Threat intelligence
- Custom detection rules — admin-defined byte-pattern rules (YARA-style subset: hex strings with
?? wildcards, severity, category, description) in Admin → Detection rules. Synced to clients via the threat-feed channel (rules_version); the Windows client evaluates them during scans after exact SHA-256 and fuzzy-hash matching, through the normal quarantine/alert pipeline. Ships empty; patterns are validated server-side.
Server / fleet admin
- Custom alert rules — "tell me when X happens": critical detection, license/trial expiring within N days, install offline for N days, quarantine retrieval completed. Rules are evaluated every 15 minutes against real DB state; firings send email (SMTP) and appear as dashboard notifications. Builder UI in Admin → Alert rules; all changes and firings are audit-logged.
- API tokens — scoped keys for third-party integrations: create/revoke in Admin → API tokens,
read scope (detections/installs/licenses) and commands scope, sent via Authorization: Bearer header. Tokens are random 64-hex strings shown once and stored hashed. - Multi-admin roles — owner (ADMIN_PASSWORD, unchanged) plus a read-only viewer role enabled by the
VIEWER_PASSWORD env var. Viewers see dashboards and reports; every mutating admin endpoint returns 403 for them (enforced server-side, with a read-only badge in the UI). - Scheduled PDF reports — the weekly email report can now attach a PDF (hand-rolled writer, no new dependencies). Toggle "Attach the report as PDF" in Admin → Notifications.
*Not tested on real Windows yet — installer UI, mic/protected-folder/banking prompts, app-limit enforcement, and the LAN scanner still need verification on a physical machine.*
Version 1.5.1
Account creation in the subscription and free-trial flows
- Subscribe page — an optional password field sits right in the checkout form: add a password (8+ characters) and your online account is created automatically during checkout. Your new license is linked to the account immediately, and the success page confirms "Your license is linked to your account". Leave it empty and checkout works exactly as before.
- Download (free trial) page — a compact account card lets you create your free account with email + password before installing; it tracks your trials, devices, and license keys. Hidden automatically when you're already signed in.
- Installer — the Windows setup wizard now asks for an optional account email; it's saved in the client config and sent with the trial/device registration so the server can link the trial to your account. Console (silent) installs accept it as a parameter too.
- Trial-to-account linking — registering a trial with an email stores it on the license; if an account with that email already exists (or is created later and signs in), the trial and all licenses for that email are linked to the account automatically.
- Privacy on checkout — entering an email that already has an account never creates a duplicate and never reveals that the account exists; checkout just proceeds normally. Short passwords are rejected with a clear "8 characters" message.
*Not tested on real Windows yet — installer UI, the new account-email field, firewall prompts, clipboard/WLAN behavior, DNS changes, and the integrity monitor still need verification on a physical machine.*
Version 1.5.0
Server / fleet admin
- Quarantine file retrieval — the client reports its quarantine list (id, name, path, size, time, threat) in each heartbeat (max 50). The admin can request any file; the client uploads it on the next heartbeat via an install-authenticated endpoint (25 MB cap, 30-day expiry, strict id validation). Download files from Admin → Installs. All request/upload/download actions are audit-logged.
- Policy templates — one-click presets: High security, Balanced, Family, Performance. Preview the exact diff before applying; applying a template is total. Templates cover every policy key, including the new DNS filtering toggle.
- Syslog / SIEM forwarding — detections can be forwarded to a syslog server (RFC-5424-style) in parallel with email/webhook. Configure host/port/proto in Admin → Settings; failures never interrupt detection ingestion. The admin shows an honest "not configured" state until a host is set.
- License analytics — Admin → Licenses now shows seat utilization, trial conversion, and license growth charts (inline SVG). Data before 1.5.0 is labeled honestly as unavailable rather than guessed.
- Maintenance windows — per-install quiet hours (days + start/end, overnight ranges supported) suppress scheduled non-critical weekly emails. Manual sends and critical detection alerts always go through. Set per install or fleet-wide in Admin.
Client protection (Windows)
- Per-app outbound firewall prompts — when an unknown program makes its first outbound connection, AEAntivirus asks allow/block and remembers the decision per executable (Settings lists them; changing a remembered decision requires the tamper PIN). Existing threat-feed IP blocking is unchanged.
- Clipboard guard — watches for clipboard hijacking that swaps a copied cryptocurrency wallet address; restores the original value and logs the attempt. Toggleable.
- Wi-Fi security check — warns when you join an open (unencrypted) network. Advisory only, one warning per network per session; open-auth networks using WEP no longer trigger it.
- DNS-level filtering — a second layer under hosts-file protection: blocks known-malicious domains at DNS resolution. Fail-open (a DNS failure never breaks browsing), fleet-manageable via the new
dnsfilter policy key, restores your original DNS on shutdown. - File integrity monitor — keeps a local SHA-256 baseline of protected system locations and alerts on unexpected changes; rebuild the baseline from Settings. AEAntivirus's own files are excluded.
*Not tested on real Windows yet — installer UI, firewall prompts, clipboard/WLAN behavior, DNS changes, and the integrity monitor still need verification on a physical machine.*
Version 1.4.9
Threat intelligence
- Fuzzy-hash variant detection — signatures.db is now v3 format (
<sha256> <fuzzy> <severity> <category> <name> [description]), fully backward compatible with v1/v2 lines. The client computes a structural hash (PE section layout + imports) during scans and catches slightly-modified malware variants via conservative similarity matching (tunable threshold, default 3 of 4 chunks). - Known-good whitelist — admin-managed list of trusted SHA-256 hashes; the client downloads it with signature updates and checks it *before* flagging — whitelisted files are never quarantined. All whitelist changes are audit-logged.
- Threat descriptions — signatures can carry a description of what the threat does; shown in detection alerts, quarantine details, and scan history.
- Scheduled threat-intel pull — the server can periodically fetch a threat-feed URL you configure in admin (URL + interval + optional credentials, which travel in HTTP headers only, never the query string) and merge new domains/IPs into the feed. Last-pull status shown in admin.
- Honesty note: the shipped database still carries only the genuine EICAR test signature. Real threat intelligence belongs in the server threat feed (bulk import or scheduled pull). AEAntivirus never ships fabricated malware hashes.
Client protection
- Webcam protection — detects processes loading camera-capture stacks, prompts allow/block on first sight, remembers per application (best-effort on Windows 7).
- Registry guard — watches classic persistence points (Run keys, Winlogon, etc.); alerts and offers to revert unexpected changes.
- Hosts file protection — guards the hosts file against external modification; offers to restore AE-managed entries.
- Boot-time scan — optional scan of autorun locations and critical system dirs shortly after boot (Settings toggle + fleet policy).
- Suspicious script alerts — flags malicious PowerShell/wscript/cscript invocation patterns (encoded commands, download cradles) at execution time; conservative patterns, toggleable.
- Application block list — fleet policy / per-install list of blocked apps; blocked apps are terminated on sight and attempts logged (ships empty).
Version 1.4.8
- Installer auto-close — the installer now closes on its own ~1.5 seconds after a successful install; on failure it stays open so the error can be read. The console installer no longer waits for Enter on success.
- Overlapping text fixed — audited every dialog: fixed a Settings overlap (startup note vs. Tamper Protection header) and a clipped exclusions hint; stat cards widened so large numbers like "1,234,567" fit without touching neighbors; stat value labels now repaint with an opaque background so old digits can't ghost.
- Live scan numbers — during a scan, FILES SCANNED / THREATS STOPPED update live (~4 Hz via thread-safe progress messages) with clean erase-and-repaint, so numbers never smear or overlap.
Version 1.4.7
- Secure file shredder — 3-pass overwrite (0x00, 0xFF, random) then delete; tray menu "Shred files…" and Explorer right-click "Shred with AEAntivirus", with a cannot-be-undone confirmation.
- Vulnerability scanner — inventories installed software and flags versions below the admin-managed safe list (Server → Software section; ships empty, no fabricated CVE data).
- Startup program manager — view and enable/disable startup entries (Run keys + Startup folder) from Settings.
- Uninstall PIN protection — the tamper-protection PIN now also gates uninstalling; without it, uninstall is refused with a clear message.
- Custom folder scheduled scans — schedule scans of your own folders (days + quick/full), alongside the weekday full/quick scans; fleet-enforceable via
policy_scan_folders. - Desktop icon option — choosing "User interface (system tray)" in the installer now offers a desktop shortcut (default on); the shortcut and the app itself show the AE logo.
- AE logo app icon — the AE cyan-hex mark is now the application icon of the Windows executables (taskbar, window, desktop shortcut).
- State-of-the-art look — restyled public site (hero, pricing cards, mobile-first) and dashboard-style admin panel with stat cards and detection-trend charts; refreshed tray window and dialogs (Segoe UI, dark/cyan).
- Dashboard charts — detection trends per day, last 30 days, inline SVG (no external CDN), in the admin overview.
- Audit log CSV export —
GET /api/admin/audit.csv. - Trial extension — one-click "extend trial 7 days" in the admin panel, audit-logged.
- Scheduled email reports — automatic weekly summary via the SMTP config (in-process scheduler; sends only while the server runs), plus a manual "send now" button.
- Web filter schedules — per-account time-based web-protection rules (e.g. stricter 21:00–07:00), synced to clients via the threat feed.
Version 1.4.6
- Scheduled scan type — weekday scheduled scans can now be Quick or Full (Settings; fleet-enforceable via
policy_scan_type with the usual stash/restore). - Pause with auto-resume — pause protection for 15/30/60 minutes from the tray menu; it turns itself back on when the timer expires (tamper PIN still gates pausing; resume is PIN-free).
- Ransomware canary files — decoy files planted in Documents/Desktop; if ransomware touches one, it fires a CRITICAL alert and the aggressive ransomware response.
- Beta update channel — per-install or fleet-wide choice of Stable/Beta (
policy_update_channel); beta clients pull manifest-beta.json. - Threat signature v2 format —
<sha256> <severity> <category> <name> (severity: low/medium/high/critical), backward compatible with the old 2-field lines. Alerts, quarantine, and scan history show severity + category; critical detections get the strongest alert treatment. - Threat feed metadata — feed entries carry severity/category/description (admin-editable; bulk import accepts
value,severity,category,description). - Email notifications — detection alerts and license-expiry warnings via SMTP (
SMTP_HOST/PORT/USER/PASS/FROM/TO; honest "not configured" state when absent). - Bulk remote commands — queue scan/update/uninstall for a whole account or hand-picked installs in one confirmed action, audit-logged.
- Expiring license alerts — admin overview warns about trials/licenses expiring within N days (default 7, configurable).
- Detections CSV export —
GET /api/admin/detections.csv, mirroring the license export. - Admin 2FA — optional TOTP authenticator codes on top of the admin password.
- Device tags — free-text tags per install (
front-desk, kids-pc), filterable in the admin panel. - Honesty note — the shipped signature database contains only the genuine EICAR test signature; real threat intel is imported via the admin bulk import. No fabricated malware hashes, ever.
Version 1.4.5
- Hardcoded server URL — the installer no longer asks for the update & monitoring server URL; it is baked in at build time (
https://aeantivirus.com). --set-server and the tray Settings field can still repoint an installed client later.
Version 1.4.4
- Scan exclusions — skip folders (
C:\Games) or file types (*.tmp) during scans: manage the list in Settings, and the fleet can push exclusions via policy (fleet-managed exclusions lock the local list, which is restored when released). - Tamper protection — optional PIN (stored as a salted SHA-256 hash, never plaintext) that guards Settings changes and pausing protection; the fleet can require a PIN on managed PCs.
- Missed-scan catch-up — if the PC was off or asleep at a scheduled scan time, the scan runs a few minutes after the next boot instead of being silently skipped.
- Right-click scanning — "Scan with AEAntivirus" in the Explorer context menu for files and folders (registered at install, removed on uninstall).
- Do-not-disturb — balloon popups are held while a fullscreen app or game is in the foreground, then shown afterward.
- Scan history — every scan is logged locally (date, duration, files scanned, threats found) and the tray status window shows the recent scans.
- Remote uninstall — retire a device from the fleet with a confirmed remote command.
- Admin audit log — every admin action (policy changes, license actions, commands, threat-feed edits…) is recorded with a timestamp in a new Audit log section.
- Bulk licenses + CSV export — generate up to 100 keys per action and export the license list to CSV.
- Seat-limit warnings — licenses at or near their device limit now show FULL / NEAR LIMIT badges.
- License reinstatement — revoked keys can be reinstated to active; protection resumes on the next heartbeat.
Version 1.4.3
- Scan frequencies — scheduled scans now support Weekly (every week), Every 2 weeks (selected weekdays on even ISO weeks), and Monthly (first occurrence of each selected weekday each month), in Settings and in the fleet policy (fleet overrides lock the local controls as before).
- Admin command center upgrades — the fleet dashboard is now a polished, professional console with a sticky section nav: per-install detail view (event timeline, stats history, detections, command history), stale/offline badges + filters, per-install scheduled-scan status (last full/quick + next scheduled run), license actions (revoke, extend expiry, change tier), bulk threat-feed import/export, install/detection search & filters, policy presets (save/apply named policy snapshots), command history, per-install policy overrides, and a detection webhook (POST new detections to your URL) with a Notifications section and test button.
- Multi-account management — the admin panel now manages each customer account separately: an Accounts section shows every account's licenses, installs, and detections, each account can have its own fleet policy (heartbeat resolution: per-install override → per-account policy → global policy), and legacy installs without an account appear under "Unassigned".
- Polished user UI — the tray status window and Settings dialog got a professional redesign: statistics cards with thousand separators, a symmetric action-button grid, sectioned Settings with a dark-themed frequency dropdown, and refined owner-drawn buttons (hover, focus, and disabled states). The dark cyan AE identity is unchanged.
Version 1.4.2
- Modern cybersecurity look — the old logo is gone. The tray UI now uses a runtime-drawn cyan hexagonal AE brand mark (tray icon included), a near-black/navy palette with cyan accents, a cyan protection-status indicator, and dark owner-drawn action buttons.
- Scheduled scans — pick the weekdays and the time (24-hour
HH:MM) for an automatic full scan; the engine checks every minute and runs once per qualifying day. Set it in Settings (all days unchecked = off), or enforce it fleet-wide from the admin dashboard's Fleet policy — fleet schedules lock the local controls, and when the fleet releases the schedule your previous local setting is restored. - Service-mode status notifier — the Windows service used to run completely silently. Now a lightweight per-user notifier (started at logon via HKLM Run, one per session, no engine of its own) shows the AEAntivirus tray icon with live scan stats in its tooltip and pops up balloon notifications: protection started, scan results, threats quarantined/deleted, and updates. Double-click the icon (or right-click → *Protection status*) for an on-demand summary.
Version 1.4.1
- Full fleet policy — the admin dashboard now enforces every protection module fleet-wide: Firewall, Web Protection, Advanced RTP, and USB guard on/off, pushed on each heartbeat. Also new: toggle lock (users can't override your settings — the status window shows "managed by your fleet administrator"), adjustable threat-feed refresh interval (1–24h), ransomware heuristic sensitivity (off / normal / aggressive), and quarantine action (quarantine or delete immediately).
- Remote actions — from the Installs table, tell any machine to scan now, update now, or reboot (60-second warning). Commands are delivered on the next heartbeat.
- Outdated-client flag — installs running an older version than the published update get an "outdated" badge in the dashboard.
Version 1.4.0
- Control Panel uninstall — AEAntivirus now appears in Settings → Apps / Control Panel → Programs and Features with a proper entry (name, version, publisher, size). Uninstalling from there (or the Start Menu's *Uninstall AEAntivirus* shortcut) stops the service, removes firewall rules and web-protection hosts entries, deletes program files, and cleans the registry. The uninstaller itself is removed at the next restart.
- Firewall — manages Windows Firewall directly: malicious IPs from your threat feed become outbound block rules automatically. Toggle it in the status window; disabling clears all AEAntivirus rules. (It orchestrates Windows Firewall — not a kernel driver.)
- Web Protection — blocks malicious domains system-wide via a marked section of the Windows hosts file, refreshed from your server's threat feed every 6 hours. Manage the blocklist in the admin dashboard's Threat feed section (domains + IPs).
- Advanced Real-Time Protection — three new behavior layers: a process guard that scans every newly started program and kills + quarantines on a signature hit; a ransomware heuristic (beta) that watches Documents for mass-modification bursts and stops the suspicious program; and a USB guard that scans removable drives on arrival. Each layer respects the license gate and the pause switch.
- New look — the tray status window is now a dark-themed dashboard: branded header, live "You're protected" status, and iOS-style toggle switches for Real-time protection, Firewall, Web Protection, and Advanced RTP, plus scan/update/quarantine/settings actions.
Honest limits: the new modules are cross-compiled and API-checked but not yet run on a real Windows PC — install, uninstall-from-Control-Panel, firewall rule creation, hosts-file blocking, and the ransomware heuristic all need real-machine testing before relying on them. The ransomware heuristic is behavioral and labeled beta; it can only see what user-mode APIs reveal.
Report a bug (tray UI): "Report a bug..." in the tray menu and status window sends the user's description plus auto-attached diagnostics (app/Windows version, license state) to your server, where it appears in the admin dashboard's Bug reports section with Resolve/Reopen.