wfx-loot: 4,837 Hijacked GitHub Accounts Pushed a CI Secret Stealer
Starting at about 12:45 UTC on 10 November 2026, someone holding working credentials for thousands of real GitHub accounts began opening pull requests titled "docs: dependency audit notes" from a branch called wfx-docs. The pull requests were probes. Where the account turned out to have write access, the attacker came back with a commit that added a credential stealer to the repository's CI, let it run once, and reverted it about a minute later. Researchers have named the campaign wfx-loot, after the artifact name in its code.
The counts below come from the three write-ups I read in full, by Endor Labs, Chainguard and OX Security, all published on 12 November. All three credit the first disclosure to Aikido Security. I have not read Aikido's post myself, so where a number traces back to it I say so.
What happened
There was no typosquatted package and no compromised action. The attacker logged in as real developers. Endor reports one victim saying a leaked GitHub CLI OAuth token was the way in, and says this is not confirmed for the campaign as a whole. Commits were authored under the victims' own noreply addresses, unsigned, which is why they looked routine in a busy repository.
| Measure | Figure | Who reported it |
|---|---|---|
| Hijacked accounts | 4,837 | Aikido, repeated by Endor, Chainguard and OX |
| Repositories probed by 11 Nov | 61,764 | Aikido, cited by Endor and Chainguard |
| Repositories that received malicious commits | 8,093 (OX, citing SafeDep) to 8,926 (Endor, 12 Nov 21:35 UTC) | Counts taken at different times |
| Pipelines where the stealer ran and exfiltrated | At least 1,806 | Endor and OX |
| Probe pull requests I found on 16 Nov | 67,527, of which 56,988 still open | My GitHub search, below |
The malicious commits came in three shapes, and the commit messages are the easiest thing to search for:
- "ci: refresh build cache" added steps to an existing workflow. Endor's example step passes
toJSON(secrets)andtoJSON(vars)into the payload as environment variables, which hands over every secret the workflow can see. - "chore: add ci helper" added a hidden
.ci-xxxxxx.shscript and apostinstallhook inpackage.json. That variant runs on the nextnpm installanywhere, including a laptop, not only in CI. - "chore: add dependency audit workflow" added
.github/workflows/build-audit-xxxxxx.ymlon a throwawaychore/branch.
The best-known victim is WebdriverIO. OX and Endor both describe two "chore: add ci helper" commits pushed on 11 November under the project's bot account, which bypassed branch protection. A contributor spotted them within about 20 minutes and a pull request reverting both commits, webdriverio/webdriverio#16444, was opened 19 minutes after they landed.
What the payload takes
Endor and Chainguard describe the same roughly 1,500-line shell script, unobfuscated, in four stages. It collects Actions secrets and variables, GITHUB_TOKEN, OIDC request tokens, cloud and developer credential files, .env and kubeconfig files and runner memory. It then uses the job's GitHub OIDC token to mint AWS, Google Cloud and Azure credentials and tries to make them last: new IAM access keys, an AWS role trust policy rewritten to "Principal": {"AWS": "*"}, Azure federated credentials named wfx-ci-.... Finally it posts a forged success status check and ships the loot to :8080/exfil on two IP addresses, with Telegram added as a second channel from 12 November according to OX.
The target list includes AI keys: OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot and Codex. That is why this is on the AI agent security incident log. After LiteLLM's 40-minute backdoor last month, model API keys are now standard loot, priced like cloud keys.
Two details matter for anyone investigating. The commit is reverted and the workflow run and its logs are deleted, so Endor calls a missing run number in your Actions history "the most reliable" sign. And the code can upload a wfx-loot-<run id> artifact, but that path was switched off. Chainguard is blunt that the absence of such artifacts "proves nothing".
What I ran
The probe pull requests are public, so I counted them with the GitHub search API at 05:15 UTC on 16 November:
for d in 10 11 12 13 14 15 16; do
echo -n "2026-11-$d "
gh api -X GET search/issues \
-f q="\"docs: dependency audit notes\" is:pr created:2026-11-$d" \
--jq '.total_count'
done
| PR created (UTC date) | Pull requests |
|---|---|
| 10 Nov | 15,062 |
| 11 Nov | 45,681 |
| 12 Nov | 6,776 |
| 13 Nov | 2 |
| 14 Nov | 2 |
| 15 Nov | 1 |
| 16 Nov (to 05:15) | 3 |
Two things stand out. The probing stopped almost completely after 12 November: eight matching pull requests in three and a half days, which may be copycats or researchers rather than the same operator. And 56,988 of the 67,527 are still open, about 84 percent, down from the roughly 90 percent the researchers reported on 12 November. Tens of thousands of maintainers have not looked yet. Search counts are approximate and the phrase search can match unrelated titles, so treat these as orders of magnitude. I could not find any statement from GitHub about the campaign on its blog, changelog or status page as of this morning.
What to do this week
- Search your organisation. Run
gh search prs "docs: dependency audit notes" --owner <org> --created ">=2026-11-10", thengit ls-remote origin 'refs/heads/wfx-docs' 'refs/heads/chore/*'on each repository. An open probe pull request from a member account means that account's credentials are in the attacker's hands, even if nothing else happened. - Search history, not the tree. The commits were reverted. Run
git log --all --since=2026-11-10 --format='%h %an %s' | grep -E 'refresh build cache|add ci helper|add dependency audit workflow|ci: restore|ci: cleanup'and look for gaps in workflow run numbers. - If you find a hit, rotate everything the workflow could read, not only the secret you think was used. Then check cloud audit logs for new access keys, trust policy changes and federated credentials since 10 November.
- Revoke and reissue the affected account's tokens, including
ghCLI tokens and SSH keys, and review its OAuth and GitHub App grants. - Make stolen CI identity worth less. Put deploy credentials behind GitHub environments with required reviewers, bind cloud and vault roles to specific branches and environments, and set
npm ci --ignore-scriptswhere you can. My GitHub Actions OIDC guide covers the trust policies, GitHub Actions security in 2026 covers workflow hardening, and today's secrets managers for CI comparison has a tested example of a role that refuses the attacker's branches.
FAQ
What is wfx-loot?
It is a credential-stealing campaign that began on 10 November 2026. The attacker used stolen credentials for 4,837 GitHub accounts to open probe pull requests, then pushed commits that ran a secret stealer in the CI of repositories where those accounts had write access. At least 1,806 pipelines ran it, according to Endor Labs and OX Security.
How do I know if my repository was hit?
Look for pull requests titled "docs: dependency audit notes", branches named wfx-docs or chore/ followed by eight characters, commits titled "ci: refresh build cache", "chore: add ci helper" or "chore: add dependency audit workflow", and missing run numbers in your Actions history from 10 November onward. The commits were usually reverted within a minute, so check git history rather than the current files.
Is the campaign still running?
The researchers called it active on 12 November. My count of new probe pull requests fell from 6,776 on 12 November to single digits a day from 13 to 16 November, which suggests the probing phase has stopped. Stolen credentials stay valid until they are rotated, so that is not the same as being safe.
Related: AI Agent Security Incident Log, The npm Worms That Hunt AI Agents and How to Stop Secrets Leaking Into Git.
Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.
I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.
Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.
Related
Google GTIG: Agents Now Turn Patches Into Exploits Unattended
Google's 17 November 2026 AI Threat Tracker reports exploit_agent, a framework that diffs patches and fires exploits with no human, LiteLLM CVE-2026-40217 used for LLMJacking, and a DPRK cluster planting backdoors in .claude/settings.json.
6 minNov 18, 2026How to Audit a Repo's Coding Agent Config for Backdoors (2026)
Step-by-step: find the hooks, helper commands, env overrides and MCP servers a repository can make Claude Code run, tested on 2.1.327. Includes a jq audit script, a CI gate, safe flags for claude -p, and managed settings that lock it down.
11 minNov 17, 2026GitHub Copilot Can Now Force MCP Servers Through Plugins
On 16 November 2026 GitHub added a managed setting that makes Copilot accept MCP servers, hooks, skills and agents only from plugins, and VS Code set February 2027 to remove its Local agent harness. What changed and what to check.
6 min