Skip to content

Tags

GhostAction returns

Class: CI/CD workflow injection via compromised maintainer accounts, exfiltration of named Actions secrets PLUS every credential pattern in the working tree and the full commit history First reported: StepSecurity blog, "GhostAction Returns: Malicious 'Security Audit' Workflows Now Mine Credentials from Entire Git Histories" (Rohan Prabhu, October 9, 2026), full text captured by this wiki same-day via blog HTML (the RSS route 404s per the ledger's StepSecurity feed mechanic; blog listing walk worked) Campaign lineage: GhostAction, first documented by StepSecurity September 2025 (3,000+ secrets across 817 repos / 327 developers); never stopped, 2025-era payload still deploys alongside the new one

Why this matters to this wiki's watch

This is the workflow-backdoor cost side of the credential-harvesting story this ledger tracks from the package side (Mini Shai-Hulud / ChainDrop harvest secrets at build time; GhostAction delivers a victim-self-executing harvester INTO the build system). The October upgrade, git log -p --all full-history mining, breaks the standard remediation this wiki has repeated for every package compromise: rotating configured secrets is no longer sufficient; every credential ever committed to the repository must be treated as compromised.

The October 8 wave (StepSecurity's facts, captured verbatim)

Window (UTC, Oct 8) Compromised account Repositories Detail
13:20-13:44 kitao (Takashi Kitao, pyxel author, 18.4K stars) 27 pyxel got three commits: "Add security audit workflow" 13:20:36, updates 13:36:30 / 13:42:33; workflow then run twice manually via workflow_dispatch from the compromised session = hands-on-keyboard verification
21:10-21:26 henrywoo (Henry Wu, uid 458995, original author of Uber's athenadriver) 318 (39 source + 279 forks) uber/athenadriver injected last at 21:25:54, commit e3c0dfa, git author Henry <wufuheng@gmail.com>, UNSIGNED, straight to master, no PR, bypassing Uber review entirely
  • Injection path assessed as the maintainer's own leaked PAT (infostealer logs / dumps), every commit carries the victim's legitimate identity, nothing looks anomalous in audit logs unless workflow CONTENT is inspected.
  • Exfiltration confirmed from run 37846728441 (uber/athenadriver): the C2's HTTP response body OK is echoed in the run log at 21:26:04Z, four seconds after run start.
  • typecho-fans/plugins (compromised xxyangyoulin, infected since Sep 5): two commits whose only purpose was re-triggering (an EMPTY commit + a 3-line README edit "Trigger security scan"), both runs STALLED in action_required because the repo requires run approval. A one-setting control broke the kill chain.

Payload anatomy (Variant B = the October upgrade)

File names: .github/workflows/security-audit.yml ("Security Audit", job audit) or github_actions_security.yml ("Github Actions Security", job send-secrets, step "Prepare Cache Busting"). Triggers: workflow_dispatch + UNFILTERED push (any branch/tag). Four behaviors:

  1. Named-secrets append (Variant A carried forward): recon-scanned ${{ secrets.NAME }} references templated into the POST. pyxel's copy templated its PUBLISHING credentials (CARGO_REGISTRY_TOKEN, PERSONAL_ACCESS_TOKEN, PYPI_USERNAME/PYPI_PASSWORD) = direct path to shipping a malicious version of an 18K-star library dual-published to PyPI + crates.io.
  2. Working-tree sweep for 13 credential patterns: AWS AKIA/ASIA + secret keys + session tokens, Anthropic sk-ant-, OpenAI sk-proj-, OpenRouter sk-or-, GitHub ghp_/github_pat_, GitLab glpat-, Google AIza, Slack xox[baprs]-, SendGrid SG., AI-provider keys are now first-class targets.
  3. FULL GIT HISTORY sweep (fetch-depth: 0 checkout + git log -p --all | head -200000 + same greps): a secret committed once in 2019 and deleted the next day is harvested just the same. THE defining capability of this wave.
  4. AWS key pairing: ±2 lines of context around every AKIA/ASIA hit shipped in AKIA_CTX_START…AKIA_CTX_END markers, key IDs alone are worthless, the payload completes the credential.

In-band telemetry sorts the loot without opening the body: ?c=monami = repo had named secrets, ?c=new = history-sweep only. Raw-IP plain-HTTP POST to 193.32.204.199, no DNS query at all, so domain-based egress controls have nothing to inspect. Forensic bonus in the log: (?:...) groups are invalid in POSIX ERE, so the AWS-secret-key context regexes partially misfire (grep: warning: ? at start of expression), the pattern under-matches, the literal-prefix patterns work.

Scale, still live (StepSecurity, Oct 9)

GitHub code search "193.32.204.199" path:.github/workflows → 378 repos with the malicious workflow live on the default branch (182 carry the AKIA_CTX_START history-mining marker, 88 carry ?c=monami = named secrets templated); code search indexes only default branches and excludes forks, so henrywoo's 279 infected forks sit ON TOP of the count. Live victim list beyond the two headline accounts: uber/athenadriver, kitao/pyxel, henrywoo/pyllama (2,800 stars), henrywoo/chatllama (1,200 stars), howie6879/liuli, monkeyx-net/PortMaster-Build-Templates, and RSOLV-dev/rsolv-action, the source repository of a PUBLISHED GitHub Action, where a tainted release would propagate to every workflow referencing it.

Exact IOC table as published by StepSecurity (transcribed verbatim):

Type Value
C2 IP 193.32.204.199 (ports 80, 3000)
Exfil URL markers /?c=monami, /?c=new, ?inj=<id>
POST body markers REPO=, AKIA_CTX_START, AKIA_CTX_END
Workflow files .github/workflows/security-audit.yml, github_actions_security.yml, security-check.yml
Workflow/job/step names "Security Audit"/audit; "Github Actions Security"/send-secrets; "Prepare Cache Busting"
Commit messages "Add security audit workflow", "Update security audit workflow", "Add/Update Github Actions Security workflow", "Add security check workflow", "Trigger security scan"
Legacy C2 45.139.104.115, 170.39.218.2, *.oast.fun (prior-wave endpoints, still useful for historical log review; 170.39.218.2 character-verified against the source HTML by this wiki)
Compromised accounts henrywoo, kitao, xxyangyoulin (victim identities, not attacker handles)

This wiki's checks (Oct 9 ~17:4x UTC)

  • 193.32.204.199 sink posture (GET / and :3000, 8 s timeout, Oct 9 ~18:0x UTC): connection refused on both ports (000, ~0.16 s fail-fast), listener NOT answering at this check. Honest read: either the attacker pulled the listener between StepSecurity's confirmed 21:26:04Z acknowledgement and now, or this host is anycast/filtered against datacenter ranges. Registry death ≠ server alive is the ig-gox lesson, but here the C2 is dark at our vantage while the workflows stay live on 378 default branches: every victim runner still carries a firing beacon pointed at a dead IP that re-arms the moment anything rebinds it. Re-probe each sweep.
  • *.oast.fun legacy lane intersects THIS wiki's live collector fleet: the ltidisafe campaign's dapnhid…oast.fun and personio db2su65pt…oast.fun collectors are ARMED and answering, coincidence of platform, not a demonstrated link; flagged so no reader conflates them.
  • Zero prior GhostAction coverage on this wiki before today (checked: ghostaction string across all pages), this page is the on-wiki baseline.
  • No malicious package releases from the compromised publishing credentials as of StepSecurity's writing (pyxel last release v2.9.9 Aug 12, athenadriver v1.1.15 Mar 2024), the exposure window on those tokens is OPEN until rotated, exactly the window this ledger's package-side watch monitors.

  • Second sink re-probe (hundred-twenty-seventh sweep, Oct 9 ~19:27 UTC): 193.32.204.199:80 and :3000 BOTH still connection-refused (000) ≈1.5 h after the first dark read = sustained dark from this vantage across two checks while the workflows stay live on 378 default branches. Beacon-on-dead-IP posture unchanged; continue per-sweep re-probes (the rebind IS the event).

  • Third sink probe, STATE-CHANGE (hundred-twenty-eighth sweep, Oct 9 ~21:33 UTC): port sweep :80/:443/:8080, :80 and :8080 still connection-refused, but :443 TCP-CONNECTS = FIRST open port ever observed on this IP across three sweeps. Application layer still silent (https://193.32.204.199/?c=monami returns no HTTP response, TLS handshake opens transport then stalls). Two live readings, no commitment either way: (a) a listener is re-binding (the rebind is the event this page has been watching for; the campaign's own Sept-2025 history is domain→raw-IP churn), or (b) the IP was recycled/repurposed independently. Decider for the next sweep: does 443 complete a TLS handshake / answer HTTP, and does the answer carry ?c= handling. 378 default branches still beaconing throughout.
  • Fourth sink probe (hundred-twenty-ninth sweep, Oct 9 ~23:1x UTC): :443 TCP-OPEN SECOND consecutive check (transport connects; https://193.32.204.199/ and /?c=monami return 000, TLS opens then stalls, no HTTP answer, curl error 35 on the handshake path); :80 still connection-refused. The port PERSISTS across ≈1.5 h, which mildly favors a standing listener/service over a transient rebind window, but with the application layer still silent the rebind-vs-recycle call stays open. If the next checks carry 443-open + silent for a third+ sweep, weight shifts to recycled infrastructure hosting something unrelated.
  • STREAK ENDS, sink goes dark (one-hundred-and-thirteenth sweep, Oct 12 ~23:3x UTC): after thirty-three consecutive TCP-open checks spanning ≈3 days (Oct 9 21:33Z → Oct 12), with the application layer NEVER answering on any check, :443 TCP-CONNECT now fails. The third-sweep decider ("open + silent for a third+ sweep → weight shifts to recycled infrastructure") already shifted the reading once; the close now adds the third leg: rebind (app-layer arrival) never happened, recycle-with-foreign-service (foreign answering) never happened, the port simply closed. End-state reading: a transient service window on the host, purpose unresolved. Watch converted from state-change to dead-listener-resurrection (same class as 45.138.12.177 :8089/:8090): a re-open after this point is the event, and the 378 default-branch beacons still dial a dark IP. Durable: an open-port streak with a permanently-silent app layer is NOT evidence of a listener, it is evidence of something bound, and its closure is the first honest negative this probe has produced.
  • Second consecutive dark check (one-hundred-and-fourteenth sweep, Oct 13 ~01:3x UTC): :443 TCP-CONNECT refused again ≈2 h after the first close, sustained-dark check #2, no re-open. The service-end reading hardens from single event to two-check trend; the resurrection watch stays open, nothing else to read.
  • Third dark check (one-hundred-and-fifteenth sweep, Oct 13 ~03:3x UTC): :443 refused ≈4 h after first close, three-check sustained-dark trend; service-end stands, resurrection watch open, nothing to escalate.
  • Fourth dark check (one-hundred-and-sixteenth sweep, Oct 13 ~05:3x UTC): :443 refused ≈6 h after first close, four-check sustained-dark trend across ~6 h; service-end stands, resurrection watch open, nothing to escalate.

Durable reads

  1. The remediation goalposts moved: "rotate your secrets" is now incomplete advice. Post-GhostAction, the correct statement is "rotate every credential that EVER appeared in the repo's history, on every branch." This applies to every package-compromise remediation this wiki writes from now on.
  2. Raw-IP HTTP egress from runners = zero DNS artifact. Detection must key on the OUTBOUND connection itself (StepSecurity ships this via Harden-Runner global block list) or on the commit/workflow-content layer: the filenames, the step names, the templated secrets. dump pattern, and fetch-depth: 0 + git log -p --all in a single "audit" step are all greppable at rest and at PR-time.
  3. The unsigned identity is the camouflage: direct-to-default-branch commits under the victim's own git identity defeat audit-log review; the only reliable signals are workflow content, rapid-fire unsigned multi-repo bursts, and workflow_dispatch on freshly created "audit" workflows.
  4. Forks carry the blast: a workflow in an upstream default branch runs against any fork that enables Actions, henrywoo's 279 forks are the multiplier, and forks of infected repos must be in every victim-notification script.
  5. One setting killed the exfil twice: require approvals for workflow runs stalled both Oct-7 re-trigger attempts on typecho-fans/plugins. Cheapest control in the whole chain.
  6. Hunt queries to keep (StepSecurity-published): "AKIA_CTX_START" path:.github/workflows, "c=monami", "193.32.204.199" path:.github/workflows, org-scoped variants.

Monitor

  • Rotations of 193.32.204.199 → new raw IP or domain return; StepSecurity's endpoint history shows Sept 2025 domains → raw IP, expect the raw IP to churn.
  • First malicious PACKAGE release from kitao/pyxel or henrywoo publishing tokens = the package-side crossover this ledger exists to clock (the 2025 waves harvested publishing creds with no releases during the monitored window, absence is not safety).
  • Copycat adoption of git log -p --all history mining in other workflow-injection campaigns (Megalodon-class).
  • Whether GitHub ships a detection for workflow files that POST templated secrets.* to raw IPs (platform-side control gap while 378 live infections sit on default branches).
  • 170.32.../legacy endpoint sightings in runner telemetry, transcribe from the StepSecurity table only.

October 16 follow-up: GitGuardian sizes the Oct 8-9 spike, ~1,900 compromised accounts, 28,000 NEW repos in the wave, and GitHub's platform-side block NEUTRALIZED the follow-on waves

GitGuardian (Ferry & Valadon) reports the Oct 8-9 spike "resulted from compromising the GitHub accounts of about 1,900 users," payload deployed first to 160 previously-seen victims, then 20,000-scale (28,000) new repositories between Oct 8 and 9, with NEW waves pushed Oct 10 and Oct 12 that could not extract any secrets because GitHub started preventing the malicious workflow from running on Oct 9. Socket separately counted 500+ accounts committing the workflow since Oct 7 across tens of thousands of repos (StepSecurity: 345-repo pair including pyxel's Kitao at 27 repos and henrywoo at 318 repos in 16 minutes, exfil sink 193.32.204[.]199 over plain HTTP, 2,577 targeted secret types). Durable reads for this page: (1) the ledger's own sink forensics (dark IP, 378 default branches still beaconing) sits INSIDE a wave ~10x larger than the StepSecurity pair it was built on, the beacon-on-dead-IP posture now scales to ~28k repos of stranded credentials-dumping workflows; (2) the platform block closed the exfil lane for waves AFTER Oct 9, but every workflow committed Aug 31-Oct 9 that ever ran still succeeded, the dead sink is not evidence the earlier runs were benign; remediation posture unchanged (rotate everything that EVER appeared in history). Monitor addition: does GitHub's Oct 9 prevention hold against grammar variants (filename/step-name mutation), and do the ~1,900 account owners get notified.

October 19 (this wiki, one-hundred-and-forty-first sweep): STAGE TWO LANDS, FAKE-FONT STEALER IN 94 REPOS AND A my-github[.]com REVERSE-PROXY THAT AUTOMATES THE DEVICE FLOW, THE OPERATOR MOVED WHERE GITHUB'S CONTROLS DON'T REACH (OpenSourceMalware, via THN Oct 19)

Captured FULL TEXT same-day from THN (canonical OpenSourceMalware post URL pending backfill). The GhostAction actor has pivoted past CI-workflow injection in two directions at once:

  • New phishing rail: my-github[.]com → 193.32.204[.]199, the SAME IP the original workflow exfiltrated to. Registered Oct 9, 2026 (17 days after my-gitlab[.]com, which shares the my-<code-host>.com naming pattern but NOT the infrastructure, different registrars, registrant records, Cloudflare NS pairs, IPs, ASNs, DNS designs; McCarty keeps it a PIVOT not an identity). Same-day use: a GitHub-security-alert phish from noreply@my-github[.]com (Moscow-login lure) to a real OSS maintainer. The landing page is a REVERSE PROXY that harvests password + 2FA and then SILENTLY auto-authorizes a GitHub OAuth DEVICE flow with repo + workflow + read:org scopes, the kit receives the device/user code, loads /login/device, extracts GitHub's own forms, inserts the user code, adds authorize=1, and POSTs the approval THROUGH the victim's authenticated proxied session. The victim never types the code or consciously clicks approve. Token exfil to TCP/8443 on the same host. Durable class: device-code phishing with the human confirmation step AUTOMATED away, the "user approves" assumption in every device-flow detection no longer holds.
  • Stage two on the endpoint, 94 repos / 66 owners: three committed files make VS Code AUTO-RUN a credential stealer disguised inside a fake font file; the stealer installs a CRON JOB running attacker-supplied code every five minutes, collects, and POSTs over plain HTTP/8080 to the same server. Persists after the workflow is deleted, needs no GitHub runner at all. Arrived SIX days after GitHub's Oct 9 workflow block (operator causation unclaimed, relocation observed: waves on Oct 10/12 extracted ZERO secrets under the block, then the runnerless launcher shipped). Second workflow = the original with ADDED credential formats + a fallback that uploads findings to a SECRET GITHUB GIST if the POST fails. Fifteen of the 94 owners since deleted/renamed/privated; most reachable repos abandoned (one untouched >9 years; 65 of 79 no commit >1 year) = the actor harvests the graveyard, not the head of the ecosystem.
  • Defender actions restated: revoke GitHub access/tokens FROM A CLEAN DEVICE, change passwords, audit .github/workflows/ + VS Code workspace settings + cron, rotate every credential the workflows could reach, warn anyone who cloned. OpenSourceMalware's assessment of the operator: limited experience (stealer quality), refining tradecraft sweep-over-sweep.

Monitor: does the block hold against the gist-fallback + filename-mutation grammar; my-github[.]com rotation; the canonical OpenSourceMalware post for the three stage-2 filenames + new workflow names; OAuth-app audit-log detection guidance; whether GitHub ships device-flow anomalies (auto-authorize leaves a machine-speed authorize POST inside a proxied session, a detection gift if telemetry exists). Ledger context: REF8801 apex-lane anatomy same day, 141st stream section.