ZeroBreach Request early access
← All Attack Surface
critical

GhostAction wave pushes credential-stealing workflows into tens of thousands of GitHub repos

Tags: github, ci-cd, github-actions, supply-chain, malware, exploited

On October 8, 2026, researchers at Socket reported a new wave of the GhostAction supply-chain campaign. Two compromised maintainer accounts, those of Takashi Kitao (author of the 18,400-star pyxel game engine) and Henry Wu (original author of Uber’s athenadriver), pushed a fake workflow named security-audit.yml into 346 repositories in minutes, and Socket has since traced the same workflow to more than 500 GitHub accounts committing it into tens of thousands of repositories since October 7.

The workflow does nothing an audit would do. It collects the repository’s GitHub Actions secrets and sweeps the working tree and the full git history for cloud, AI, and SaaS credentials, including AWS keys, Anthropic, OpenAI, and OpenRouter API keys, plus GitHub and GitLab tokens, then posts them in cleartext to a hardcoded IP address.

Socket confirmed the workflow actually executed in affected repositories it checked, and in the clearest case, pyxel’s own publishing credentials were targeted by a successful run. This is a supply-chain identity failure: malicious automation arriving under a trusted maintainer’s name looks like ordinary development activity.

What to do

  • Search your GitHub organizations and accounts for commits adding .github/workflows/security-audit.yml or github_actions_security.yml since October 7, 2026, especially on dormant or forked repositories.
  • Rotate any secret that could have been exposed: CI/CD secrets, cloud provider keys, AI API keys, npm and DockerHub tokens, and tokens that appear anywhere in the repo or its git history.
  • Lock down repository and workflow permissions, require review for workflow changes, and audit maintainer accounts for leaked personal access tokens, the likely entry point.

A secret committed once lives forever in git history. ZeroBreach’s recurring server and application scans exist for exactly this class of drift: known-bad patterns, quietly reappearing, flagged on a schedule instead of after the damage.