Guide · Supply-chain security · as of 16 November 2026

A secret stealer ran in more than 1,800 GitHub repositories, and it wanted your AI keys. Four checks.

From 10 November 2026, someone used the stolen logins of 4,837 real developers to slip a secret stealer into GitHub repositories. The security firm Aikido, which found it, counts more than 1,800 repositories where it ran and sent credentials out. It calls itself wfx. Its shopping list includes the usual cloud and npm keys, and also the keys for AI services and AI coding tools: OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot and more. This guide works from Aikido’s and Endor Labs’ write-ups and ends with four checks.

How it works

A real developer’s name, a commit that lasts a minute.

It starts with a harmless pull request titled “docs: dependency audit notes”, which tests whether the stolen account can write to the repository. Then it pushes the stealer under an ordinary message such as “ci: refresh build cache” or “chore: add ci helper”, lets GitHub Actions run it, and reverts the change within a minute. The workflow run and its logs are deleted. So a scan of today’s code finds nothing. Aikido says the more reliable sign is a gap: a missing run number in the workflow history next to the attacker’s commit. How the logins were stolen is not confirmed. One victim reported that a leaked GitHub CLI login token was used.

The npm attacks behind the firewall

What it takes

Every secret the workflow can see, and more.

All of a repository’s Actions secrets and variables, the workflow’s own GitHub token, cloud logins for AWS, Azure and Google Cloud, npm and PyPI tokens, SSH keys, .env files, shell histories, and the keys of a long list of AI services and coding tools. It also reads tokens out of the memory of the runner process. On a self-hosted runner, a machine you run yourself, it collects more: cloud instance credentials, Docker state and system services. Some versions try to stay in your cloud account after the theft, so rotating keys alone may not be enough. At least 50 of the stolen accounts can publish packages, which is why the researchers expect a second wave.

Four checks for any business with code on GitHub

Ask whoever runs your repositories.

  • 1. Look for the fingerprints. Search your organisation for pull requests titled “docs: dependency audit notes” since 10 November, branches named wfx-docs, and commits saying “ci: refresh build cache”, “chore: add ci helper” or “chore: add dependency audit workflow”. Test: all three searches come back empty, and you checked that each one actually ran.
  • 2. Look for the gap. On any repository where you found a fingerprint, compare the workflow run numbers around that commit. Test: no number is missing. A missing one means assume the stealer ran and every secret that workflow could see is gone.
  • 3. If it ran, rotate everything it could reach, AI keys included. Repository and organisation secrets, cloud keys, npm and PyPI tokens, and the keys for OpenAI, Anthropic and the rest. Then look in your cloud account for access keys or trust rules created since 10 November. Test: each old key no longer works.
  • 4. Shrink what a stolen login can do. Block direct pushes to your main branch, require review for changes to workflow files, set the workflow token to read-only, and prefer short-lived, narrow tokens to long-lived ones. Test: a push straight to main by an ordinary account is refused.

What we checked on our own

We build with AI agents on GitHub all day. So we looked.

On 16 November 2026 we searched all 47 repositories on our GitHub account. We found no probe pull requests, no branches with the campaign’s names, and among 475 commits made since 9 November on our main branches, none with its messages. On our development machine there were no hidden helper scripts or install hooks, and the logs of our 7 self-hosted runners held no trace of the addresses the stealer sends to. That covers main branches and branch names, not every commit on every branch, so it is a floor, not proof. It also showed us something to tighten: our own GitHub command-line login is a long-lived token allowed to change workflows, the same kind one victim says was used against them.

AI agents for Dubai businesses

Questions

Questions about the wfx stealer.

  • What is the wfx GitHub Actions secret stealer?

    A script, about 1,500 lines long, that attackers pushed into GitHub repositories using stolen developer logins from 10 November 2026. It runs inside GitHub Actions, collects every secret it can reach and sends them out, then the attacker reverts the change.

  • How do I know if my repository was hit?

    Search for a pull request titled “docs: dependency audit notes”, a branch called wfx-docs, or commits saying “ci: refresh build cache”, “chore: add ci helper” or “chore: add dependency audit workflow” since 10 November. If you find one, check the workflow run numbers around it for a gap.

  • Does it steal AI API keys?

    Yes. Aikido lists keys for OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot, Codex, Ollama and other AI tools among what it collects, next to cloud, npm and SSH credentials.

  • Is this the same as GhostAction?

    No. Aikido assesses them as separate campaigns by different actors, although accounts hit by GhostAction in October were much more likely to be hit again.

  • We found nothing. Are we safe?

    Safer, not certain. The commit is reverted within a minute and the logs deleted, so check run numbers as well as code. And make check four a habit: it limits what the next stolen login can do.

Sources

read 16 November 2026

Check it at the source.

  1. Aikido Security: compromised GitHub accounts push CI secret stealer (12 November 2026) (the original disclosure: scale, what it steals, the run-number sign)
  2. Endor Labs: the wfx campaign in detail (12 November, updated 13 November 2026) (commit messages, file and branch names, and hunting commands)
  3. WebdriverIO: the revert of the malicious commits (one well-known project caught within about 20 minutes)
  4. GitHub Docs: security hardening for GitHub Actions (token permissions, environments and workflow review)

A question?

Code on GitHub?

Tell us how your repositories and pipelines are set up. We will run these four checks with you and leave you with the rules that make a stolen login less useful.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com