Threat intel
wfx-loot pushes a CI secret stealer into 61,764 repos
The wfx-loot campaign used stolen GitHub logins to probe tens of thousands of repos, then ran a payload that steals CI secrets and climbs into the cloud.
On November 10 a threat actor sitting on a pile of stolen GitHub credentials began opening pull requests across tens of thousands of repositories, one after another, to find out which logins still worked. Aikido, which named the campaign wfx-loot after the attacker's own labels for the malware and its loot, counted 4,837 compromised accounts opening these probe pull requests in 61,764 repositories. In at least 1,835 of them a follow-up payload ran and sent the repository's secrets to a server the attacker controls.
If anyone who can write to your repositories has had a token or session stolen, this reaches you. The probe shows up as a harmless pull request titled "docs: dependency audit notes" from a branch called wfx-docs, and one of those alone is enough to treat the account as compromised. What comes next reads the repository's continuous integration (CI) secrets, climbs into any connected cloud account, and rewrites trust settings so that rotating the stolen secrets later does not lock the attacker out.
How it works
GitHub Actions runs automation jobs on GitHub's own servers when an event fires in a repository, such as a push or a pull request. Those jobs usually need credentials: a token to publish a package, a key to deploy to a cloud account. The repository keeps them as Actions secrets, and a workflow receives one only when it names it, as ${{ secrets.NAME }}. A stolen login with write access can add a workflow that names every secret and does whatever it wants with the values.
The campaign splits the job in two. The probe comes first. Opening a pull request costs nothing and confirms the stolen credential still has write access without dropping anything noisy, which is why the "docs: dependency audit notes" pull request from a wfx-docs branch is the opening move. Aikido found 10,431 of these probes went to repositories the account did not own, reached through collaborator or organization membership.
The payload arrives three ways. It injects about 1,500 lines into an existing workflow, or it adds a script together with an npm postinstall hook in package.json, or it adds a fresh workflow on a short-lived chore/ branch with eight random characters. The attacker reverts the commit within a minute and deletes the workflow run and its logs, so a gap where a run number should sit can be the only trace left behind.
The script itself works in four stages. It collects first: GitHub secrets, variables and runtime tokens, cloud and developer credential files, environment variables, keys for AI providers such as OpenAI, Anthropic, Cursor and Copilot, shell histories, and process memory on the runner. Then it escalates. The prize here is OpenID Connect (OIDC) federation. Many teams have stopped storing long-lived cloud keys in GitHub and instead let a workflow request a short-lived token at run time: the job presents an OIDC token that proves which repository and branch it is, and the cloud account, trusting GitHub's issuer, hands back credentials for a named role. wfx-loot walks that path. From inside a workflow it exchanges those tokens to reach AWS, Google Cloud, Azure, HashiCorp Vault and Kubernetes, creates long-lived AWS identity and access management (IAM) access keys, and impersonates service accounts.
Persistence is where it gets patient. The script forges "success" status checks so branch protection waves its commits through. In the cloud it rewrites an AWS role's trust policy to accept any AWS account, and adds federated identity credentials named wfx-ci-... to Azure applications. Those changes survive a secret rotation. You can reset every token the workflow touched and the attacker still holds a way back in. The collected material leaves as a gzipped archive by HTTP POST to 194.150.166[.]244:8080/exfil or 18.143.168[.]96:8080/exfil. Because the destination is a raw address over plain HTTP, the job makes no DNS lookup, so DNS filtering never sees it.
What attackers are doing
The run started around 12:45 UTC on November 10 and moved quickly. The 4,837 compromised accounts opened probe pull requests across 61,764 repositories. Aikido puts the count of repositories where the payload actually ran and credentials left at 1,835, though its own indicators section lists 1,806, and it notes the true figure is likely higher because the attacker deletes its logs.
At least 50 of the compromised users publish packages, on npm, Go, Packagist, PyPI or crates. That is how a theft of CI secrets reaches past GitHub into the software supply chain, since a stolen publishing token lets the attacker ship code under a trusted name.
Two package releases are confirmed to carry the campaign through automated publishing. On Packagist, digitalelvis/neuronai-studio 3.14.3 is affected, and any machine that installed or opened it should be treated as compromised. In Go, github.com/onuragtas/docker-env tags v1.0.278 and v1.0.280 are flagged, though there the payload sits only in a workflow file that Go builds never run, so the risk is lower.
This is not GhostAction, the October campaign that committed fake "security" workflows to steal CI secrets, even though the two look alike. Aikido assesses wfx-loot as the work of a different actor. The overlap is real, though: accounts GhostAction hit were far more likely to be hit again, and in more than 100 repositories wfx-loot was injected straight into GhostAction's leftover workflow files.
What to do
- Hunt across every repository anyone could write to since November 10. Look for a
wfx-docsbranch, achore/branch with eight random characters, aci-check-NNNNbranch, a pull request titled "docs: dependency audit notes", and commits such as "ci: refresh build cache" or "chore: add dependency audit workflow". Check for a.ci-xxxxxx.shfile or apostinstallhook added topackage.json, and for a missing workflow run number, since a deleted run leaves a gap in the sequence. - Treat a probe on its own as account compromise. If the "docs: dependency audit notes" pull request appeared, the stolen credential worked, whether or not you can find the payload.
- Rotate if the payload ran. Reset the repository's Actions secrets, every cloud credential the workflow could reach, and the runtime tokens it would have collected.
- Audit the cloud for the persistence steps, because rotation alone does not clear them. Look for new IAM access keys you did not create, AWS role trust policies that name a
"*"principal or any account, and Azure federated identity credentials namedwfx-ci-.... Confirm that required status checks on protected branches actually ran rather than being marked passed. - Handle the packages. Remove or avoid
digitalelvis/neuronai-studio3.14.3 andgithub.com/onuragtas/docker-envv1.0.278 and v1.0.280, and treat any machine that installed the Packagist package as compromised. - Block 194.150.166[.]244 and 18.143.168[.]96 on port 8080 at the egress firewall, and search outbound logs from your runners back to November 10 for either address.
- Revoke and reissue the GitHub tokens of any account that opened a probe, sign those users out everywhere, and review who holds collaborator and organization write access while you are there.
OIDC federation moved the valuable thing out of the secret store and into the trust policy. Rotating a key is visible work. A trust policy that quietly accepts any account, or an extra federated credential on an Azure app, is not, and it is what lets an attacker return after the clean-up looks finished.
We harden GitHub organizations, cloud trust policies and Kubernetes access as part of our cloud and Kubernetes security work, and review CI/CD pipelines, token scopes and workflow permissions in our red team and code review work. For help checking your repositories or planning a rotation, open the chat and Yaali, our AI agent, will pass your question to the engineer who would run it.
Sources: Aikido, Help Net Security.