Skip to content
nulltapE-reader edition

← All e-reader articles

5 min read

WFX stealer erases GitHub Actions runs after stealing CI secrets

Stolen write tokens let attackers alter CI and delete runs. Hunt commit history and cloud federation after a run disappears; later Telegram uploads bypass the published collection IPs.

Read on the standard site

Editorial illustration of a missing workflow run slot in a CI history beside a remaining run card.

A stolen GitHub write token can turn a routine CI run into a secret-stealing job, then erase that run from the Actions tab. StepSecurity’s November 12 investigation documents the WFX campaign doing this to repositories it could write to, including an executed stealer in a surviving Nextcloud workflow run. The attack started November 10 and was still changing delivery methods on November 12.

The researchers counted more than 63,000 search results for the attacker’s probe pull requests at 07:40 UTC on November 12. The count measures access probes; stealer execution depends on a subsequent payload push and run. A repository owner who finds one should investigate the writing identity and the commits and runs that followed; where the injected job ran, assume the credentials available to that runner were exposed.

A harmless pull request tests access

The attacker opens a pull request titled docs: dependency audit notes from wfx-docs, adding docs/wfx-audit.md. The PR is a probe for a token that can write to a repository. The subsequent payload push depends on that token’s permissions and the repository’s CI setup. StepSecurity found more than 21,000 commits titled ci: refresh build cache between November 10 and 12; correlate matching commits with Actions runs and treat a matching missing run as probable execution when scoping exposed credentials.

The attacker used valid developer and bot tokens to push changes through normal repository write access. StepSecurity says 173 probe PRs from 19 accounts were performed via the Fly.io GitHub App, identified in GitHub’s performed_via_github_app field as fly-io with app ID 34557; other campaign activity used credentials such as personal access tokens. The method used to obtain those tokens remains unknown. Review affected accounts’ connected app grants and token activity; StepSecurity reports no evidence of a Fly.io infrastructure breach.

The stealer hides inside existing CI

After probing, the attacker pushes a roughly 1,500-line shell stealer to a workflow, a new throwaway-branch workflow, a hidden .ci-*.sh script invoked by a package.json hook, or an ordinary shell script already run by CI. In Nextcloud’s nextcloud/docker repository, the stealer was appended to update.sh and ran in workflow run #2675. The published log includes memscan_files 3 and archive_bytes 8971; StepSecurity interprets these as a runner-memory read and an 8,971-byte collection archive. A green job can still be the run to investigate.

In the workflow variant, one step passes secrets and variables to the next; other variants read the Runner.Worker process memory to get credentials the job can access even if the workflow file never explicitly names them. The script also reads environment variables and cloud credential files, and attempts to use the job’s OIDC cloud sign-in to reach AWS, GCP and Azure. StepSecurity found attempted persistence through AWS role trust changes involving Principal and AWS wildcard values, new IAM access keys, and Azure app federated credentials named wfx-ci-*. Its check-marking capability was switched off in all 101 copies it examined. A developer who pulled the transient package.json hook and installed dependencies could also have executed the stealer locally; include those developer credentials in the response when that happened.

Reverts and deleted runs hide execution

The attacker pushes a revert titled ci: restore, sometimes followed by ci: cleanup, then deletes the workflow run and logs. In drunomics/nuxtjs-drupal-ce, StepSecurity records a stealer commit followed 12 seconds later by a bot-account revert; the Actions tab skips run #1964 between #1963 and #1965. As the researchers put it, “A missing run means the stealer probably ran and the attacker deleted the evidence.” Correlate the gap with the commit and account before scoping exposed credentials.

Three collection IPs appeared in recovered scripts: hxxp://194[.]150[.]166[.]244:8080/exfil, hxxp://18[.]143[.]168[.]96:8080/exfil and hxxp://153[.]75[.]82[.]159:8080/exfil. From about 05:30 UTC November 12, sampled new copies sent their archives to Telegram through the legitimate shared service api.telegram.org, without a collection IP; every copy StepSecurity sampled from 07:00 to 07:40 UTC used that route. Blocking the three IPs alone therefore leaves the later delivery path open. An egress allowlist for build jobs that excludes bare collection IPs and api.telegram.org covers both observed network routes; check existing job dependencies before enforcing it. The researchers say the Nextcloud run’s first upload to the first IP succeeded, based on the run’s log behavior.

Recover the identities behind each run

Search pull requests since November 10 for docs: dependency audit notes and branch wfx-docs; inspect commit history for ci: refresh build cache, chore: add ci helper, chore: add dependency audit workflow, ci: restore and ci: cleanup. Examine workflow files, package.json hooks and scripts CI invokes, including their reverted versions. Match suspicious commits to Actions run numbers and GitHub audit records for deleted runs. A probe calls for token investigation; a matching injected run calls for incident response.

Revoke the affected account or app token and remove malicious CI changes. For a job that ran, rotate repository, environment and organization secrets it could access, plus credentials on a self-hosted runner; review its OIDC role grants and the cloud sessions it could start. Check AWS role trust for the reported Principal/AWS wildcard combination and new IAM access keys since November 10; correlate CloudTrail AssumeRoleWithWebIdentity, AssumeRole, CreateAccessKey and UpdateAssumeRolePolicy with wfx-* sessions; check Azure app federated credentials for wfx-ci-*; inspect GCP GenerateAccessToken and SignJwt events from the GitHub workload identity pool. Remove unauthorized trust and access keys before calling key rotation complete.

For ongoing runs, GitHub’s security guidance says anyone with repository write access can read its configured secrets, so narrow write-capable identities and protect privileged workflows. GitHub’s token documentation describes setting permissions at workflow or job level to give GITHUB_TOKEN only the access each job needs. That limits one credential’s reach; it cannot undo a malicious run already executed with a separate stolen write token. Reconcile the commit sequence, run sequence and cloud federation changes after containment; those three histories show whether a clean current workflow actually corresponds to a clean credential path.

Primary sources