Trapdex Subscribe

CampaignsAI plugin / MCP / skill

hackerbot-claw: self-described AI agent exploiting GitHub Actions and poisoning CLAUDE.md

Do this now

  1. Remove pull_request_target workflows that check out untrusted code; sanitize branch and file names used in scripts
  2. Make AI review workflows read CLAUDE.md and similar context files from the base branch only, with least-privilege tokens
  3. Block the listed domains and rotate tokens exposed to fork-PR workflows

Indicators to check

Look for any of these on your machines or in your logs:

  • repositorygithub[.]com/hackerbot-claw

CSVJSONAll 3 indicators

What happened

A GitHub account describing itself as an autonomous agent attacked CI workflows in several large open-source projects using pwn-request, branch-name, filename and script injection. It got code execution in several repos and stole a write-scoped token from one. In another repo it replaced CLAUDE.md with injected instructions aimed at an AI review workflow, which the model flagged and refused.

  • Claude Code
  • GitHub Actions

Attack chain

  1. Lure

    Malicious pull requests and a poisoned agent context file (CLAUDE.md)

    First seen

  2. Artifact

    Crafted PRs and branch names; modified CLAUDE.md

  3. Actor

    No source we cite attributes this activity to an actor.

  4. Infrastructure

    • Payload hosting and exfil domain

    Last seen

Indicators

CSVJSON

All values are defanged (how to read them).

TypeValueCopy
domainhackmoltrepeat[.]com
domainrecv[.]hackmoltrepeat[.]com
repositorygithub[.]com/hackerbot-claw

Confidence: Moderate

Strong evidence, not yet confirmed by our own analysis. See the confidence scale for what each level requires. The sources below are the evidence.

Sources

Notes

  • Mostly CI exploitation rather than social engineering; included for the agent-context poisoning attempt.

Corrections

No corrections.

See the corrections policy. Spotted an error? Tell us.