Secretus logo
newswfx-lootgithub-actionsaikido

wfx-loot: Rotating GitHub Actions Secrets Can Leave Cloud Access Open

Aikido reports CI secret theft in the wfx-loot campaign. Check payload execution, cloud trust and runner safety before handing over replacement credentials.

·6 min read·Secretus Editorial

A replacement deployment key is only useful if the attacker has lost the ability to collect its replacement. Aikido's 12 November report on the wfx-loot campaign describes stolen GitHub accounts being used to plant a credential stealer in CI workflows. Its analysis also identifies capabilities for changing cloud trust and creating additional access. That makes recovery a question of who can run the next job and obtain cloud credentials, as well as which secrets to rotate.

A probe pull request and a stolen secret are different findings

Aikido reports 4,837 compromised accounts probing 61,764 repositories from 10 November, with credential theft in more than 1,800 repositories. The affected-repository total differs between sections of its report, so the exact figure should be treated as provisional. BleepingComputer's separate reporting corroborates the campaign's mechanics; it does not independently reproduce the researchers' count.

The probe used a branch named wfx-docs and a pull request titled docs: dependency audit notes. A probe alone does not establish that a workflow executed or secrets were exfiltrated. Aikido describes payload delivery through changed workflows, temporary branches and an npm install hook, followed by rapid reverts and deleted runs. A clean current branch therefore does not settle what ran earlier.

This is a new campaign, although some accounts overlap with the October GhostAction activity. Aikido assesses the two campaigns as distinct. The available reporting does not establish where every stolen GitHub token originated or the full private-repository impact.

Recover the permission to mint credentials

The payload analysis describes capabilities to create longer-lived credentials and alter cloud identity or trust settings where permissions allow. Capabilities are not proof that every affected account experienced every action. They are a reason to inspect cloud-side changes before declaring a rotation complete.

Think of a deployment job that normally exchanges its GitHub identity for temporary cloud access. Deleting a stored API key does not remove that trust relationship. If an attacker still controls an accepted workflow or identity, the next request can obtain fresh access. If the attacker created a separate credential, replacing the original one leaves that second credential untouched.

GitHub's OIDC documentation explains that cloud providers validate token claims against the role's configured trust conditions. Short-lived tokens reduce the need to keep long-lived cloud keys in repository secrets. The job that receives them still needs carefully bounded permissions and a trusted execution path.

A recovery checklist for the repository and cloud owners

  1. Stop untrusted execution. Pause affected workflows and restrict the compromised account's access while preserving available audit evidence. Review workflow history and install hooks, rather than relying only on the current diff.
  2. Separate what happened from what could happen. Record repositories probed, payload execution found and credentials available to those jobs. When execution is confirmed, treat that job's accessible secrets and cloud roles as compromised.
  3. Inspect cloud changes. Compare role trust, federated identities, access keys and permissions with the approved configuration. Investigate additions and unexpected grants. Assign a cloud owner to remove unauthorized access.
  4. Contain existing sessions. Use the provider's session controls as well as key revocation. AWS documents a mechanism to deny older role sessions; it affects legitimate users too and does not prevent later role assumptions. Close the path to new unauthorized sessions.
  5. Restore a trusted runner and workflow. Review the code that will receive replacements. Investigate or rebuild affected self-hosted runners through your incident process before allowing them to hold new credentials.
  6. Replace and test. Rotate each exposed credential at the service that accepts it, configure the approved consumers, and verify that the old value is rejected. Record owners, identifiers and test results without recording the values.

Do not send the new secrets into the old workflow

Keep a small recovery ledger: credential name, accepting service, intended consumer, owner, replacement time and verification result. A single shared cloud key often turns out to have several consumers. Finding those owners before rotation helps avoid an outage being mistaken for another intrusion.

GitHub recommends least privilege and protected environment secrets, which can require a reviewer before a job receives them. It also warns that automatic log redaction is not guaranteed. Use those controls when restoring CI, and review whether a job needs the secret at all. A build that only compiles public code should have a clear reason for receiving production access.

When a human must receive a replacement, verify the recipient through an established contact route and send only the value needed for that task. Keep the status update in incident chat; put the credential in the approved runtime store. Avoid pasting a complete environment file into a ticket to explain which single variable changed.

Where Secretus helps with the handover

Secretus can carry a replacement text secret between verified people with browser-side encryption and a link that becomes invalid after expiry or first successful access. Use trusted devices for that transfer. It does not revoke cloud sessions, repair trust policies or protect a value after a compromised runner reads it. Those controls must be ready before the handover, and the credential still needs its own expiry or revocation at the destination.

Sources