4,837 Hijacked GitHub Accounts, 61,764 Repositories Probed: The wfx-loot Campaign Is a CI Secret Apocalypse in Slow Motion

4,837 Hijacked GitHub Accounts, 61,764 Repositories Probed: The wfx-loot Campaign Is a CI Secret Apocalypse in Slow Motion

4,837 Hijacked GitHub Accounts, 61,764 Repositories Probed: The wfx-loot Campaign Is a CI Secret Apocalypse in Slow Motion

Starting around 12:45 UTC on November 10, an attacker armed with stolen credentials for 4,837 GitHub accounts began opening pull requests titled "docs: dependency audit notes" from branches named wfx-docs against 61,764 repositories — a write-permission probe at unprecedented scale. Where the probe found write access, the follow-up arrived: a roughly 1,500-line shell script researchers at Aikido have named wfx-loot, delivered by modifying an existing GitHub Actions workflow, adding a new workflow on a short-lived branch, or planting an npm postinstall hook. By November 12, researchers at SafeDep counted 8,093 repositories that received malicious commits on default branches alone — the true total is higher — and the stealer demonstrably executed in at least 1,806 of them, with secrets confirmed stolen from over 1,835 targets.

The script's ambition is what sets this apart from routine token scraping. It harvests GitHub Actions secrets and variables, GITHUB_TOKEN, OIDC and runtime tokens; developer and infrastructure credentials from Git, SSH, GPG, Docker, Kubernetes, npm, PyPI, Vault, and Terraform; and environment variables, .env files, private keys, keystores, shell histories, process memory, and workflow artifacts. It then goes lateral: it exchanges GitHub OIDC tokens for temporary AWS, Azure, and GCP sessions, chains into additional IAM roles to mint long-lived AWS credentials, and impersonates GCP service accounts. In 771 repositories the payload arrived as an npm hook — meaning it executed not just in CI but on developer machines during npm install. And roughly 50 of the compromised accounts hold publishing rights on npm, PyPI, Go, crates, and Packagist, which turned the credential theft into a supply-chain infection risk: several fake packages were pushed before the campaign was caught.

Two details should worry anyone running CI. First, Aikido found that 92% of the probe pull requests remained open days later — the legitimate owners never noticed a foreign "dependency audit" PR sitting in their queue, and every one of those PRs is a tripwire that fired. Second, the initial access was plain credential theft at scale, not a GitHub breach — 4,837 accounts had tokens or passwords weak enough to be harvested, which is why the researcher's recovery list is unglamorous: revoke the commit author's tokens, OAuth grants, SSH keys, and sessions; rotate every secret the repository can read; enable 2FA; and treat any machine that ran npm install on an affected commit as compromised.

What Should You Do?

  1. Search your org for wfx-docs branches and "docs: dependency audit notes" pull requests — an open PR from that campaign means an account with write access was already compromised, whatever the payload did next.
  2. Where the payload executed, treat it as a total credential compromise: rotate Actions secrets and variables, GITHUB_TOKEN trust boundaries, cloud OIDC trust policies, package-registry tokens, and SSH/GPG keys — the script's collection list is the rotation checklist.
  3. Audit cloud OIDC trust. The script converts GitHub OIDC tokens into AWS/Azure/GCP sessions, so restrict the sub condition in your cloud trust policies to specific repositories and environments — a wildcard subject condition is an open door.
  4. Anyone who ran npm install on an affected commit should rotate that machine's keys — the postinstall hook variant runs on developer laptops, not just in CI.

The WAF Angle

wfx-loot never touches a WAF, and that's exactly the point: your build pipeline is an application with internet-reachable authentication, and its tokens are the new session cookies. A CI system trusts whatever presents a valid token, exactly as a web app trusts a session cookie — and this campaign is the pipeline equivalent of 4,800 stolen session cookies used to log in, enumerate, and escalate. Defenders should apply the same discipline web teams learned years ago: least-privilege token scoping (per-environment, per-repo, short-lived), monitoring for anomalous token use (a workflow running at 3 a.m. from an unexpected runner is the CI version of impossible travel), and blast-radius reduction — a secret readable by one workflow should not be readable by all of them. The campaign also echoes GitLab's KEV'd CI/CD flaw still running on 1,000+ exposed servers and the trojanized Claude Code extension: the pipeline layer keeps producing the same class of incident — trusted code, untrusted provenance — and the fix is provenance, not perimeter.

Sources