A credential-stealing campaign used 4,837 compromised GitHub accounts to test access to 61,764 repositories, then pushed malicious code into projects where the stolen credentials had write permission. Security firm Aikido, which named the campaign wfx-loot, found evidence that the payload ran and exfiltrated credentials from at least 1,835 repositories.
The activity began at about 12:45 UTC on November 10. It was not a conventional malicious pull request that required a maintainer to merge code. The attacker first opened a harmless-looking probe, then used the compromised account’s existing write access to modify workflows or package-install behavior directly. In many cases, the malicious commit was reverted and its GitHub Actions run deleted within a minute.
Teams that find the indicators should not stop after deleting the pull request or rotating repository secrets. Aikido’s analysis shows that the payload could exchange GitHub OIDC tokens for cloud sessions, create durable AWS credentials, and add Azure federated identities. Those changes can preserve access after the original GitHub token and CI secrets have been replaced.
First, determine which of three events occurred
The campaign generated several large numbers that describe different stages of the attack. Treating them as interchangeable can lead to either unnecessary panic or an incomplete response.
- Probe only: A pull request titled
docs: dependency audit notesfrom a branch namedwfx-docsconfirmed that a stolen account could reach the repository. A probe means the GitHub account or token was compromised, even if no malicious workflow ran. - Malicious commit: The attacker injected code into an existing workflow, added an npm
postinstallhook, or created a temporary workflow. Separate research cited by the Cloud Security Alliance found malicious commits in 8,093 public repositories. - Payload execution: Aikido found evidence of execution and credential exfiltration in at least 1,835 repositories. This is the highest-confidence compromise group, but deleted run logs and activity in private repositories mean the total may be higher.
Search repository and pull-request history from November 10 onward for the branch wfx-docs; short-lived branches matching chore/ followed by eight characters; and commits titled ci: refresh build cache, chore: add ci helper, chore: add dependency audit workflow, ci: restore, or ci: cleanup. Missing run numbers associated with those commits are important because the attacker routinely deleted workflow runs after exfiltration.

What the stealer could collect
The roughly 1,500-line shell payload searched far beyond values explicitly passed into a workflow. It collected GitHub Actions secrets and variables, the job’s GITHUB_TOKEN, OIDC and Actions runtime tokens, the full environment, and recent workflow artifacts. It also scanned process memory and common credential paths for AWS, Azure, Google Cloud, Docker, Kubernetes, npm, PyPI, SSH, GPG, HashiCorp Vault and Terraform.
AI development credentials were part of the target list. The script looked for keys and configuration associated with OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot, Codex, Continue, Codeium, Aider, Weights & Biases and Ollama. It recursively searched the runner’s home directory for files such as .env, .pem, .p12, .jks, .kdbx and Kubernetes configuration files, and captured shell histories for several common tools.
Self-hosted runners deserve a wider investigation. On those systems, the payload also attempted to read cloud instance-metadata credentials, Docker state, runner RSA material and systemd units. If an npm postinstall variant ran on a developer workstation or persistent runner, rebuild or forensically examine that host rather than assuming a repository cleanup removed the risk.
Why GitHub secret rotation is not enough
The payload could request GitHub OIDC tokens for AWS, Azure and Google Cloud audiences and exchange them for cloud sessions. That is important because OIDC is normally used to avoid storing long-lived cloud keys in GitHub. In this incident, a compromised workflow could use the same federation path during its brief execution window.
In AWS, the script attempted to assume additional roles, create access keys for IAM users, and rewrite a role trust policy with a wildcard principal. In Azure, it could use Microsoft Graph access to add a federated identity credential named with a wfx-ci- prefix to an application. That new relationship points back to the compromised repository, allowing a later workflow to authenticate without a stored secret.
Responders should therefore inspect cloud control-plane logs and current identity configuration, not just access-key inventories. In AWS, review role-assumption activity, CreateAccessKey calls and trust-policy changes during and after the repository’s exposure window. In Microsoft Entra ID and Azure, review application credential changes, unfamiliar federated identity credentials, service-principal activity, Key Vault access and new token exchanges. In Google Cloud, inspect service-account impersonation, token generation and IAM changes.
GitHub’s OIDC guidance recommends conditions that restrict which repositories, branches and workflows can request cloud access. Those conditions reduce blast radius, but they must be verified after an incident because wfx-loot specifically tried to alter the trust relationship on the cloud side.
A practical wfx-loot response sequence
- Contain the GitHub identity. Revoke the suspected personal access token, OAuth authorization, GitHub CLI credential or other session. Review the affected user’s security activity and every organization and repository the identity could write to.
- Preserve evidence before cleanup. Export repository history, pull requests, workflow metadata and organization audit logs. GitHub’s organization audit log can be filtered by actor, repository, action and time, and exported as JSON or CSV. Do not rely on Actions logs alone because the campaign deleted them.
- Inventory the workflow’s reachable credentials. Include organization, repository and environment secrets; OIDC roles; package registries; deployment systems; signing keys; AI provider keys; and any files or metadata available to a self-hosted runner.
- Remove persistence before rotating credentials. Check AWS trust policies and IAM access keys, Azure application federated credentials, Google Cloud service-account permissions, Vault roles and Kubernetes service accounts. Otherwise, new credentials may be exposed through access the attacker already preserved.
- Rotate and revoke by blast radius. Start with credentials that can mint other credentials or change identity policy, then cloud and package-publishing access, deployment keys, GitHub tokens, AI API keys and lower-impact service secrets. Review provider billing and usage logs for stolen AI and cloud keys.
- Investigate runners and developer machines. Rebuild ephemeral runners from known-good images. For persistent or self-hosted systems, examine shell history, systemd changes, cached cloud sessions, SSH material and unexpected files before returning them to service.
- Review releases and packages. Aikido identified
digitalelvis/neuronai-studioversion 3.14.3 on Packagist as malicious. It also flagged workflow payloads ingithub.com/onuragtas/docker-envversions 1.0.278 and 1.0.280; the Go source and release binaries were unchanged, but the affected tags should still be removed from builds.
Controls that limit the next stolen-token campaign
Organizations should replace broad classic personal access tokens with fine-grained tokens or GitHub Apps where possible, keep base repository permission at the minimum level, and regularly remove stale collaborator access. Changes under .github/workflows/, package lifecycle hooks and organization rules should require focused review through rulesets or CODEOWNERS.
Workflow permissions should be declared explicitly, with contents: read as the normal baseline and id-token: write limited to jobs that actually need federation. Cloud trust policies should bind access to named organizations, repositories, branches, environments or reusable workflows instead of accepting a broad GitHub subject. Protected deployment environments can add approval and branch restrictions before a job receives sensitive access.
The wfx-loot campaign succeeded because a stolen developer identity connected source control, CI execution and cloud authority. Recovery has to follow the same chain. A clean-looking default branch is not proof of a clean environment when the workflow ran, exported credentials, changed cloud identity, reverted itself and removed its own log.