Start with the uncomfortable part. Nothing was hacked. No zero-day, no CVE, no clever memory trick. Attackers authenticated with valid stolen GitHub tokens and used GitHub Actions exactly the way it was designed to be used.
The campaign has two names. Aikido calls it wfx-loot. StepSecurity calls it WFX Phantom. Same script, same exfiltration infrastructure, same "wfx" string the malware writes into its own logs. Activity started around 12:45 UTC on November 10 and was still spreading when both reports went public on November 12.
4,837 stolen accounts. 61,764 repositories touched. 21,000+ payload commits. Roughly 1,800 repos that actually executed the stealer. 92% of those pull requests still sitting open. That was two days of work.
Three steps, all of them legitimate
The loop is almost insultingly simple, because it does not need to bypass a control. It only needs to avoid human attention.
- Probe. Open a pull request titled
docs: dependency audit notesfrom a branch namedwfx-docs, adding a two-line markdown file. The file is byte-identical every time. Its only job is to prove the token can write. - Push. Commit the payload under a message that reads like plumbing, most often
ci: refresh build cache. - Clean. Revert within a minute, then delete the workflow run and its logs.
That third step is why so many owners are still unaware. On a typical repository the only trace is a missing run number. Seeing #1963 followed by #1965 means #1964 existed and someone deleted it.
Four places the payload hid inside GitHub Actions
All of it is the same script. Only the delivery changed to suit the repo.
- Inside a trusted workflow. Two extra steps bolted onto an existing
ci.ymlortests.yml. This is the dominant variant, with more than 21,000 commits. - Inside a shell script the workflow already runs. Appended to files like
update.shortests/retry_env.sh. - As a brand new workflow.
.github/workflows/build-audit-xxxxxx.ymlon a throwaway branch. - As an npm install hook. A hidden
.ci-xxxxxx.shplus apackage.jsontweak, usuallypostinstall.
Notice what the dominant variant deliberately avoids: planting a fresh workflow file. It edits a workflow the repository already trusts, on the default branch, or appends to a script that workflow already executes. GitHub's July 2026 protection that holds "potentially malicious" workflow runs for approval is genuinely effective against the noisy version of this attack. Against this one it has very little to see.
Confirmed victims include nextcloud/docker, webdriverio/webdriverio (56 commits across 28 repos, pushed by its own bot account), and kitao/pyxel. Maintainers have since reverted most of those commits. Two Go modules and a Packagist package shipped with the payload attached, which is how a repository problem turns into someone else's install-time problem. We saw the same shape in the cPGuard update compromise, where a trusted update channel did the delivery work for the attacker.
What GitHub Actions credentials theft actually grabs
The script is about 1,500 lines of shell, and it does not politely collect the secrets your workflow mentions. It dumps ${{ toJSON(secrets) }} and ${{ toJSON(vars) }} in one move, then reads the memory of the runner process itself through /proc/<pid>/mem. Every secret that job could touch, decrypted, whether or not the YAML ever referenced it.
Then the usual tour: AWS, Azure and GCP credential files, Docker config, kubeconfig, Vault and Terraform state, SSH and GPG keys, npm and PyPI and Packagist publish tokens, shell history, every .env in the home directory. Plus AI provider keys, which is the detail that should make any team experimenting with copilots sit up straight. OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot, and a long tail of others.
This is not a repository compromise. It is a credentials compromise that happens to announce itself through a repository.
You can see the whole thing in one place, because one run survived. In nextcloud/docker, run #2675 was never deleted, and its public log shows the script describing its own work: memory files scanned, GCP auth empty, archive bytes 8,971, and no retry line, meaning the first upload succeeded. The job finished green. Nothing looked wrong, because nothing was wrong by any rule the runner enforces.
Rotating keys will not close this
This is the part I would put in front of every engineering lead. The script installs two back doors that survive credential rotation entirely.
- In AWS, it rewrites the trust policy of a reachable role to
"Principal": {"AWS": "*"}. That role is then assumable by literally any AWS account. - In Azure, it adds federated identity credentials named
wfx-ci-*to app registrations, trusting workflows in the compromised repo. Those workflows authenticate to Azure afterwards with no stored secret at all.
Change every password you like. Those two doors stay open. GitGuardian watched the same pattern play out through the October GhostAction wave, where 1,382 of 1,792 pushing accounts were already known from an earlier round and still worked. Teams rotated and skipped cleanup, so they got hit twice.
And GitHub? No advisory, no changelog entry, no confirmation of this campaign. Silence is not evidence of inaction, but it does mean the burden of detection sits with you.
What to check this week, in order
Order matters here. Detect, then rotate, then clean, then prevent. Rotating before cleanup gets you reinfected.
- Hunt for the gap in your run numbers. Open the Actions tab on every repo with recent activity and look for the missing run right after a strange commit. Check Settings, then Logs, then the audit log for deleted workflow runs and who deleted them.
- Grep for artifacts. Search your workflows for
toJSON(secrets),WF_CALLBACK,Restore cache, anddocs/wfx-audit.md. Look for stray.ci-*.shfiles and branches namedwfx-docsorchore/xxxxxxxx. - Rotate wide, not narrow. Everything that job could reach, including publish tokens, SSH keys and AI provider keys. Then check your registries for any version published between the bad commit and its revert.
- Kill the persistence. In AWS, find role trust policies containing
"AWS": "*"and CloudTrail events for session names startingwfx-. In Azure, delete every federated identity credential namedwfx-ci-*. - Require review on
.github/workflows. A CODEOWNERS rule on that path plus branch protection. This is the single highest-value control you can add, because it interrupts the exact step the attacker needs. Do not exempt bot accounts. The webdriverio commits came from that repo's own bot. - Cut the exits. An egress allowlist on runners kills both the HTTP channel and the Telegram channel the operator switched to mid-campaign. A bare IP on port 8080 has no business in a build pipeline.
Two smaller ones, because they cost almost nothing: use npm install --ignore-scripts in CI so postinstall hooks cannot fire, and stop approving held workflow runs blindly. If a run is sitting at action_required on a branch you do not recognize, that is probably the attack, not a stuck job.
The unglamorous half of the fix
Ten thousand of those probe pull requests landed in repositories the compromised account did not own. One stale contractor token, one forgotten outside collaborator, exposes every project that person could write to. Auditing collaborators and removing the ones nobody recognizes is unglamorous and it is half the fix. If you want a sense of how much of this sits around unnoticed, one scan of Gitee turned up 26,400 working secrets in public projects.
So no, this was not a GitHub vulnerability, and waiting for a patch means waiting forever. GitHub Actions credentials theft on this scale is a revocation problem, a permissions problem, and a review problem, all three of which were sitting in plain sight before November 10.
If you are looking at your own Actions setup right now and not enjoying the view, that is a reasonable reaction. This campaign touched tens of thousands of repositories with nothing but tokens people forgot to revoke.
Book a 15-minute CI/CD exposure check
We will look at your workflow scope, secret inheritance and linked cloud roles, and tell you what is exposed. No obligation, no sales deck.