@veai-ru/ai-agent (npm), a malicious-by-design terminal AI coding AGENT
⚠️ RETRACTION NOTICE (October 14, 2026, 128th sweep): THE SOURCE DISCLOSURE HAS BEEN WITHDRAWN. Queue PR
#1617, the sole maliciousness claim behind this page, was CLOSED UNMERGED2026-10-14T06:59:32Zafter the Veai publisher rebutted the claims claim-by-claim (06:38:02Z) and the filer conceded in full: "I treated Veai's prompt customization, usage telemetry, MCP configuration loading, and account login as evidence of malicious behavior. On re-review, I couldn't substantiate the claims that it sends full conversations or MCP configurations to the telemetry endpoint, or that the OAuth flow is harvesting credentials… I'm sorry for the incorrect report." This wiki's independent verification covered ONLY: the obfuscator.io class ofdist/extensions/*.js, the plaintext endpoint strings baked in the tarball, and endpoint liveness (plugin.veai.ru/telemetryanswering405to GET,app.veai.ru200, Yandex bucket403-exists), the exfil/theft SEMANTICS were reported "per the PR" and were never independently decoded. A reachable telemetry endpoint plus unreadable code is suspicion, not proof. Treat every malice claim below as DISPUTED until decode-level or dynamic evidence appears either way. Zero OSV, zero GHSA, package still live and publishing (0.4.0/0.5.0dogfood rungs since the filing). 128th sweep retraction analysis.
Tags
- tools
- supply chain attack
- npm
- AI agent
- AI coding agent
- system prompt injection
- session exfiltration
- telemetry exfil
- MCP configuration theft
- OAuth credential harvesting
- Keycloak
- PKCE
- obfuscator.io
- Yandex Cloud
- veai.ru
- explyt
- curated lane
- ossf malicious-packages
- zero advisory coverage
- Russia-nexus infrastructure
What it is
Curated queue PR ossf/malicious-packages #1617 (filed 2026-10-10 03:40:36Z, finder credit smiling-hyena, the same named finder as the ledger's Sep-21 radio-player-theme jsDelivr-routed PoC entry) files npm @veai-ru/ai-agent, "Veai AI Agent: a terminal AI coding agent." This wiki's same-day verification at ~05:3x UTC:
- LIVE at
latest = 0.3.2(published 2026-10-09 09:54:13Z). Version ladder from the registrytimeblock:0.0.0-stageOct 8 09:38:17Z →0.0.0-bootstrap.009:46:31Z →0.3.2-dogfood.1Oct 9 09:17:36Z →0.3.209:54:13Z, the0.0.0-stagestaging-stub grammar the ledger already tracks on attacker tooling. - ZERO OSV + ZERO GHSA at query; the record path so far is queue PR only (unmerged), same-day-unadvised while 209 downloads/week already indexed.
- Publisher account
explyt-ownerchatgpt2@explyt.com; a package.jsonexplyt.nodefield and adist/extensions/explyt-account.jsfile carry theexplytbrand through the artifact; the user-doc endpoint refuses anonymous reads so the account's other packages could not be enumerated; a scope-member search surfaced no sibling@veai-ru/*names. - Tarball (this wiki pull, sha256
280a2ce8ff3bbf8f4891e0f03d25dadbb3f754cfeeb6870bd4f1cdca5b572c72, ~6.4 MB unpacked): a genuinely complete-looking terminal coding-agent product, realREADME,agent-docs/,THIRD-PARTY-NOTICES, MCP adapter, subagent definitions (hawkeye/retriever/workhorse), vendored typebox/zod/yaml, with the hostile surface indist/extensions/*.js, whichpackage.jsonauto-loads via the pi-package mechanism"pi": { "extensions": ["./dist/extensions/*.js"] }. The README states the agent "runs with exactly the extensions this build ships with", the extensions ARE the product surface, and they are where the hostile behavior lives.
Mechanics (static inspection of the published tarball, no execution)
What PR #1617 asserts, checked file-by-file against the shipped bytes:
- System-prompt replacement (
dist/extensions/identity.js, 33 KB). The file is a textbook obfuscator.io string-array-rotation blob (verified header shape:parseIntchecksum loop + rotated string table + hex indexes, the exact family the ledger already profiles onchai-loggerand the PhantomRaven-adjacent sets). No plaintext survives to read; per the PR the decoded array embeds fragments of well-known AI-agent harness system prompts used as injection templates that replace the active agent's system prompt at session start. This wiki's confirmation is the obfuscation itself + the file's session-start hook role, not a decoded read. - Full session exfiltration (
dist/extensions/telemetry.js, 68 KB). Same obfuscated class; plaintext identifiers that survived includetelemetryandconversationId. The PR enumerates the hooked event set,agent.ToolCall.Started/Completed,tool_execution_start/end,session.Session.Started,agent.Context.Compacted, and payload fieldsconversationId, workspaceId, model, inputTokens, outputTokens, username, hostname, platform, all POSTed tohttps://plugin.veai.ru/telemetry. - MCP configuration theft (
dist/extensions/mcp.js, 139 KB). Largest extension; zero plaintext config-path hits in this wiki's grep (.claude/settings,mcp.json,.mcp.json,mcp_servers.json), consistent with the same string-array obfuscation; per the PR it reads and exfiltrates those MCP config files. The target list is exactly the MCP-config inventory an agent-supply-chain researcher would expect. - OAuth credential harvesting (
dist/extensions/explyt-account.js, 19.5 KB). This wiki confirmed plaintextcode_verifierandkeycloakidentifiers; per the PR it drives an OAuth PKCE flow againsthttps://app.veai.ru/keycloakand stores harvested access tokens in the OS keychain via@napi-rs/keyring(the package even ships a standalonedist/mcp-keyring-helper.cjskeychain read/write/remove helper, clear text, taking JSON requests on stdin). - Binary delivery via Yandex Cloud (
dist/cli/brand-profile.js,dist/install/installer.js). CONFIRMED IN PLAINTEXT:https://storage.yandexcloud.net/veai-releases/agent/(bucket answered403, exists, listing denied),https://app.veai.ru/,https://app.veai.ru/keycloak,https://app.veai.ru/api/v1/data,https://plugin.veai.ru/telemetry,https://plugin.veai.ru/openai/api/v1, the last one meaning the package can also route the victim's model traffic through the operator's proxy. The installer additionally bootstraps its own Node runtime from nodejs.org and an npm root, i.e. the install path is self-sufficient.
Network state at this wiki's check (~05:4x UTC Oct 10): plugin.veai.ru/telemetry 405 (listener live, POST-only), app.veai.ru 200, storage.yandexcloud.net/veai-releases/ 403. All three C2 surfaces alive within ~20 h of the 0.3.2 publish.
Why this class matters
This is not a hijacked legit package and not a typosquat: it is a complete, self-branded AI coding agent whose business model appears to BE the telemetry, the durable defensive reads:
- The victim installs it voluntarily,
logins into it voluntarily, and types real work into it. Every credential-injection and session-tap pattern the ledger has catalogued against compromised agents (469-location collectors, MCP config reads, workflow-log theft) arrives here as advertised product behavior behind obfuscation, with a jurisdictional infrastructure signature (veai.rudomains, Yandex Cloudru-central1bucket, Keycloak IdP). - The prompt-replacement hook makes the package an attack amplifier against every OTHER agent on the same machine: the system prompt is the trust root for tool-calling, so an attacker-controlled prompt can steer the agent to exfiltrate more, approve more, or write malicious code the user then ships.
@veai-ru/ai-agentis the pi-package ecosystem (peer deps@earendil-works/pi-coding-agent0.84.4), the extensions auto-load mechanism (pi.extensionsglob in package.json) is a new delivery rail worth a grep: any pi-package with obfuscated files underdist/extensions/is the same shape.
Hunt keys
- Domains/URLs:
plugin.veai.ru/telemetry,plugin.veai.ru/openai/api/v1,app.veai.ru/keycloak,app.veai.ru/api/v1/data,storage.yandexcloud.net/veai-releases/agent/. - Files:
dist/extensions/identity.js/telemetry.js/mcp.js/explyt-account.jsobfuscator.io headers;mcp-keyring-helper.cjsinvoked with stdin JSON;~/.pior equivalent extension dirs containing@veai-ru/ai-agent. - Account: publisher
explyt-owner/chatgpt2@explyt.com; tarball sha256280a2ce8ff3bbf8f…(0.3.2). - Egress: if your org never needs
*.veai.ruor Yandex Cloud object storage, deny it, the Unit 42 negative-egress-baseline control applies directly.
State and monitors
At check: LIVE, installable, zero advisories on either lane, PR #1617 open unmerged (queue freeze pattern, #1612/#1613/#1615 also still open since Oct 9). Monitor: whether the GHSA lane mints anything for this name (post-arsya rule: walk the type=malware listing, the ID-walk is blind); whether npm enforces (deletion vs security-holder, see the kmf 12-second family wipe precedent); 0.3.x iteration or re-registration under new names; explyt brand footprint beyond this scope (the app.veai.ru/api/v1/data "Neo plugin" rail the README mentions implies a managed-backend product surface beyond npm); whether smiling-hyena's next find lands; whether pi-package ecosystem ships an extension-signaling control.
Oct 10 one-hundred-and-thirty-third sweep tick (~09:4x UTC): STILL LIVE latest 0.3.2, zero OSV + zero GHSA ≈30 h; app.veai.ru 200; plugin.veai.ru/telemetry GET 405 / POST 400, POST-only endpoint confirmed (the 405-vs-400 split says the route accepts POST and rejected the empty body; the rail is UP); Yandex bucket root 403-exists, agent/ path 404; PR #1617 still open unmerged.
Oct 11 one-hundred-and-forty-fifth sweep tick (~09:3x UTC): C2 DARK TICK #2 (~2 h deep, spanning the 144th→145th). plugin.veai.ru/telemetry 000 on repeated retries at a STABLE resolve 158.160.61.112, TCP-443 to that IP CLOSED/FILTERED, tcp-80 timeout; app.veai.ru (DIFFERENT host, 178.154.228.84) also 000; external-reachability control 200 = genuine two-host darkening, not client-side. DNS STILL resolves both names to the SAME recorded IPs = service/session down, NOT a DNS-level takedown. First dark since the 131st-139th's live reads (telemetry was 405-GET, app 200). Per the wave-5 duty-cycle durable (dark windows are session duty-cycle, never infra absence) this is logged as a RE-ARM WATCH, not an infra-loss claim; both IPs stay on the IoC list. Package unchanged: LIVE latest 0.3.2 (modified 2026-10-09T11:27:59Z), ZERO OSV + ZERO GHSA ≈46 h, PR #1617 open unmerged ~30 h. Monitor: dark duration vs the fleet's hours-scale re-arm clock; if the dark crosses ~24 h, the "operator moved C2" reading strengthens and the Yandex delivery rail becomes the re-probe priority.
Oct 11 one-hundred-and-forty-sixth sweep tick (~11:3x UTC): RE-ARM, AND THE DARK WINDOW CONTAINED AN IP MIGRATION. plugin.veai.ru/telemetry back to 405-GET (POST-only rail live again) and app.veai.ru 200 = dark tick #2 CLOSED at ≈4 h, exactly on the wave-5 hours-scale re-arm precedent the duty-cycle durable predicts. The re-arm is not continuity: both names now resolve to a SINGLE IP 135.106.213.120, vs the dark-window pair 158.160.61.112 (plugin) / 178.154.228.84 (app). The 144th/145th read "DNS unchanged = service down, not takedown" is now SUPERSEDED as sufficient: dark windows can contain re-provisioning, and re-armed endpoints at NEW addresses = re-platforming (two hosts consolidated onto one). DURABLE: hunt/block keys must carry the domains plus the UNION of all observed IPs (158.160.61.112, 178.154.228.84, 135.106.213.120), never the current resolve alone; egress policy should deny *.veai.ru by name so address churn costs the operator nothing but your blind spot. Package unchanged: LIVE latest 0.3.2 (modified 2026-10-09T11:27:59Z), ZERO OSV + ZERO GHSA ≈48 h, PR #1617 open unmerged ≈43 h. Monitor: 135.106.213.120 stability + whether the two-host consolidation persists; next dark window's resolves (a fourth IP would confirm rotation cadence); Yandex delivery rail as the re-probe priority if a dark crosses 24 h.
Oct 11 one-hundred-and-forty-seventh sweep tick (~13:2x UTC): CONSOLIDATED-IP STABILITY CHECK #2 PASSES, steady armed state on 135.106.213.120. Both plugin.veai.ru and app.veai.ru STILL resolve the single migrated 135.106.213.120 (no fourth IP, no split-back); telemetry 405-GET, app 200 = ~2 h continuous armed since the 146th's re-arm. The re-provisioned two-onto-one topology is holding, so the consolidation read stands. Package unchanged: LIVE latest 0.3.2 (modified 2026-10-09T11:27:59Z), ZERO OSV + ZERO GHSA ≈50 h, PR #1617 open unmerged ≈45 h. Monitor unchanged: stability #3 next sweep; next dark window's resolves; #1617 merge-or-rot.
Oct 13 one-hundred-and-sixteenth sweep tick (~05:4x UTC): THE dogfood CHANNEL, TWO NEW BUILDS SHIPPED WHILE THE LEDGER WATCHED latest. Registry doc modified 2026-10-12T16:44:38Z reveals two publishes this stability watch never saw because both live UNDERNEATH latest: 0.4.0-dogfood.2 2026-10-12T12:05:32Z and 0.5.0-dogfood.1 16:44:37Z, under a NEW dist-tag dogfood, full tag set now {bootstrap: 0.0.0-bootstrap.0, latest: 0.3.2, dogfood: 0.5.0-dogfood.1} = formal release-engineering channelling (bootstrap/dogfood/latest) imported into a malicious-by-design package; the operator runs a product org's version grammar. Both new tarballs pulled and statically inspected (no execution): 0.5.0-dogfood.1 2,053,349 B sha256 0667ca127b917b4e…, 0.4.0-dogfood.2 2,042,212 B sha256 3abde4b45e16f35d… (vs 0.3.2 1,622,205 B = +431 KB/+27 %). ALL baked rails survive verbatim across the bumps: plugin.veai.ru /telemetry + /openai/api/v1 (victim model-traffic proxy), app.veai.ru /keycloak + /api/v1/data, storage.yandexcloud.net/veai-releases/agent/ = infrastructure continuity through iteration. URL-diff 0.3.2→0.5.0-dogfood.1 ADDS a broad search/MCP provider surface, mcp.tavily.com/mcp/, api.perplexity.ai/search, api.z.ai + open.bigmodel.cn web_search_prime MCP endpoints, api.keenable.ai, api.ceramic.ai, api.typesafe.ai, ydc-index.io, openrouter.ai, agent-plugins.org/schemas/1.0.0/{plugin,mcp}.schema.json, and REMOVES nothing: tool-surface growth on top of intact exfil rails (honest-record: static URL/string diff; full deob of the new bundle deferred). Rails armed at same check: telemetry 405-GET (POST-only rail), app 200, both names STILL resolve consolidated 135.106.213.120 = IP-stability holds ≈2 days since the 146th's re-arm. Zero OSV + zero GHSA ≈92 h (doc created Oct 8 09:38:17Z); PR #1617 open unmerged ≈109 h while the subject iterates twice. DURABLE SELF-CAPTURE: this page logged latest 0.3.2 stability checks #1-#14; every one was true of the TAG and false of the PACKAGE, a dist-tags.latest-based clean read misses pre-release-channel attacker code entirely. Hunt/monitor the registry time block DELTA and the full tag SET, not the latest pointer; npm install @veai-ru/ai-agent@dogfood is an install path that bypasses every latest-keyed scanner. The same blind spot is symmetric for defenders using latest-only SBOMs. Monitor updated: next dogfood build (cadence Oct 9 0.3.2-dogfood.1→0.3.2 same-day, Oct 12 0.4.0-dogfood.2→0.5.0-dogfood.1 4 h 39 m apart = active iteration), 0.5.0-dogfood.1 deob, whether dogfood graduates to latest, 135.106.213.120 persistence.
Related pages
- py2ops PyPI operator-linked companion shell, same-48h, same zero-advisory curated-lane class; different target (keystrokes at a fake REPL vs full agent sessions)
- GhostAction GitHub Actions workflow injection, the CI-side mirror of the same credential-mining appetite
- Unit 42 Web3 C2 evolution on the ChainDrop page, the negative-egress-baseline control cited above
- Mini Shai-Hulud npm/PyPI worm campaign, the 469-location collector whose AI-tool config targets this package formalizes as product