ReleaseLock learn

Was my repo hit by wfx-loot, and can the stolen publishing tokens swap my releases?

By Cassie Crossley · Published November 13, 2026 · Sources checked November 13, 2026 (UTC+2)

Since November 10, 2026, a campaign called wfx-loot has used stolen GitHub credentials to push a self-erasing CI secret stealer into repositories. Check your repos for the probe pull request and for gaps in workflow run numbers. No GitHub release-asset swap has been reported in this campaign. Stolen publishing tokens are a risk to releases, not something the researchers cited here have shown happening.

How to tell if a repo was hit

Aikido, OX Security and GitGuardian describe the same pattern. Search every repository you can write to, on or after November 10, for these signs:

Commits are made as the real user of each repository, using their GitHub noreply email and no signature. OX Security says this points to stolen access tokens rather than a flaw in GitHub itself.

A run-number gap proves a run was created. GitGuardian cautions that it does not prove the run executed.

What happened, and why the npm hook mattered

GitGuardian reports a small first phase from October 24 to 29 (80 payload patches, 7 accounts, 12 repositories), a quiet stretch, then a mass wave starting November 10 at 12:45 UTC that was still running on November 12. Aikido gives the same November 10 start time.

The same stealer script ships in several wrappers. GitGuardian describes four: a throwaway workflow, an injection into an existing workflow, an npm lifecycle hook (generally postinstall), and an append to an existing shell script.

GitHub held the workflow-file forms. GitGuardian says every November payload run it could still see (7 of 7) had the conclusion action_required and never started. Those are GitGuardian's observations. We found no official GitHub advisory, blog post or changelog entry about wfx-loot as of November 13, 2026, so this page does not attribute the hold to GitHub itself.

The npm hook got past the hold. It changes no workflow file. The repository's own, unchanged workflows run npm install, npm ci or pnpm install, and that runs the hidden script. GitGuardian reports public job logs from a sample of 19 repositories showing the script ran, with a successful exfiltration reported in 18. It also counts hook commits rising sharply once the workflow forms were held: 838 in the November 11 window from 09:00 to 12:00 UTC, and 1,210 in the November 12 window from 06:00 to 09:00 UTC. Because the hook also runs outside CI, a developer who installed dependencies in the short window before cleanup could have run it on their own machine.

The numbers, by source

These researchers counted different things at different times with different methods. We list each separately and do not add them together.

SourceWhat it reported
Aikido, November 124,837 compromised accounts opened probe pull requests in 61,764 repositories. GitHub search showed 61,839 such pull requests from November 10, 12:45 UTC to November 11, 15:00 UTC. 10,431 went into repositories the account does not own, and 92% were still open. Its summary says at least 1,835 repositories ran the payload and had credentials exfiltrated. Later in the same article it says indications the payload ran in at least 1,806 repositories, and that this may understate the impact because logs were deleted.
OX Security, November 124,837 hijacked accounts. Over 66,200 probe pull requests, roughly 58,900 still open. 8,093 public repositories received a malicious commit on the default branch, 771 of them through npm postinstall hooks. Confirmed exfiltration from at least 1,806 repositories. 10,431 probe pull requests hit repositories the hijacked account does not own, including 28 WebdriverIO repositories reached through the project's bot account. At least 50 accounts publish to npm, Go, Packagist, PyPI or crates.io. More than 4,100 new probe pull requests on November 12 alone. OX says private repositories and non-default branches are not in these public counts.
GitGuardian, November 13GitHub search on November 12 returned 66,001 probe pull requests with the probe title since November 10, and 28,148 ci: refresh build cache commits on default branches between November 10 and 12. Its own dataset holds 28,500 payload patches from 2,367 accounts across 17,715 repositories. That is a different sample, not a correction of the other two.

The sources do not give one agreed total of victim repositories, and neither do we. All three describe the campaign as still running when they published. OX gives one example: two chore: add ci helper commits pushed through the WebdriverIO bot account on November 11 at 12:11 UTC, bypassing branch protection, which a contributor spotted and reverted in about 20 minutes.

What the stealer takes

This is what the script is able to collect, not a claim that every target lost every item. According to Aikido, OX Security and GitGuardian it reaches:

It can also mint GitHub OIDC tokens and try to leave cloud access behind. Aikido describes AWS role trust policies rewritten to open the role to any AWS account, and Azure federated credentials named wfx-ci- followed by more characters. Rotating secrets does not remove access like that.

Exfiltration changed during the wave. Defenders were first given HTTP addresses to block. Aikido and OX list 194[.]150[.]166[.]244:8080 and 18[.]143[.]168[.]96:8080, both at the path /exfil. Later copies use Telegram. OX dates that shift to about 05:30 UTC on November 12. GitGuardian dates Telegram delivery from its v3 payload at 01:07 UTC on November 12, and Telegram-only from 01:48 UTC. The two timestamps come from different sources and we do not reconcile them. Blocking the known addresses alone is not enough after the shift.

What leaked publishing tokens can do to releases

None of the researchers cited here reports a confirmed GitHub release-asset swap, which would mean replacing the files on an existing release. We are not claiming one. This section describes a risk path.

What has been reported is new package versions published by automated triggers. That is a different thing from a swapped release asset:

Aikido assesses the package releases as a byproduct of automated publishing. Its read is that the main goal was stealing CI secrets.

Now the risk, which is ReleaseLock's reading and not a finding of the researchers. About 50 of the compromised accounts publish to package registries, and the script goes after npm, PyPI, crates.io and Packagist tokens. A stolen registry token can publish a malicious version. A stolen GitHub token that can create releases or upload release assets could be used to replace or add files on a release. Our GhostAction guide explains which tokens can change a release, and the same applies here.

What to do if the payload may have run:

Immutable releases are the backstop if someone later uses a stolen token against a release. Once an immutable release is published, GitHub blocks changes to its assets and its tag. See Can this GitHub release be swapped? and what 2026's incidents show about unlocked releases.

What other controls would and would not have done

More background: GhostAction and stolen GitHub tokens, LiteLLM: clean tag, poisoned release and re-enabled compromised Actions. To check whether the tools you install have locked releases, search the ReleaseLock directory.

On the log side of this campaign, see what GitHub's new workflow-log secret scanning catches and misses, and why a deleted run is not proof that nothing leaked.

Quick answers

How do I know if my repo was hit by wfx-loot?

Look since November 10, 2026 for a pull request titled "docs: dependency audit notes" from a wfx-docs branch, commits titled "ci: refresh build cache", "chore: add ci helper" or "chore: add dependency audit workflow", a revert titled "ci: restore" or "ci: cleanup", a root .ci-xxxxxx.sh file with a package.json hook, and gaps in workflow run numbers. A probe pull request alone means the account that opened it is compromised.

Did GitHub block the malicious workflows?

GitGuardian observed that every November payload run it could still see in workflow-file form (7 of 7) was held with the conclusion action_required and never started. The npm hook form changes no workflow file, so the hold did not apply to it. No official GitHub statement on wfx-loot had been found as of November 13, 2026, so this comes from GitGuardian, not from GitHub.

Were any GitHub releases swapped in this campaign?

No researcher cited on this page reports a confirmed GitHub release-asset swap. Reported package releases, such as a Packagist version and Go module tags, came from automated publishing. Stolen publishing and release-create tokens remain a risk path.

If I only use a few Actions secrets, do I still rotate the rest?

Yes, if a workflow form of the payload ran. GitGuardian says it reads toJSON(secrets) and toJSON(vars), which return every repository and organization secret and variable, not only the ones the workflow references.

Where do I check release-integrity controls?

Search the ReleaseLock directory to see whether the tools you install have locked releases, and read the linked learn pages on immutable releases, stolen tokens, npm provenance and workflow lockfiles.

Sources