Wiz researchers Rami McCarthy and Hila Ramati report an ongoing campaign that uses stolen GitHub tokens to plant secret-stealing workflows. Their October 25 investigation traces Gheist activity back to September 15.

Wiz recovered more than 13,500 planted branches across more than 850 compromised accounts. Those are its observed campaign artifacts and accounts, not a count of repositories whose secrets were successfully stolen.

Private repositories lack the public approval hold

The attacker needs a compromised token with enough repository and workflow access to plant the malicious file. The workflow collects available secrets through its own job logs, downloads those logs through GitHub’s API, then deletes the run.

Encoding the secrets defeats masking that only recognizes their original values. Restricting a job’s GITHUB_TOKEN, its built-in repository credential, separately leaves the workflow’s access to configured secrets unresolved.

GitHub’s July 28 protection holds potentially malicious workflows until a collaborator approves them. GitHub says this applies only to public repositories on github.com, excluding private repositories and GitHub Enterprise Server.

Wiz says every public Gheist run it checked was held before execution. In private repositories and enterprise environments, it observed completed and deleted runs. Cloud audit records also showed workflows on attacker branches assuming AWS roles.

Exploitation status: source-reported successful activity in private environments. Wiz also observed subsequent AWS Secrets Manager enumeration with credentials from an affected organization. GitHub independently documents the approval control’s scope, but does not corroborate Wiz’s victim figures in that notice.

Deleted runs can leave evidence in cloud logs

The report gives defenders several concrete leads:

  • Repository changes: branches named ci-sync- followed by six characters, adding .github/workflows/sync-check.yml with the commit message ci: sync config.
  • GitHub audit events: runs on those branches that complete and disappear within a minute, and API activity using the user agent gheist.
  • Cloud authentication records: role assumptions or token exchanges naming a ci-sync- branch. These can survive deletion of the corresponding GitHub run.
  • Self-hosted runners: an added job named sync-host, the marker gdmp-pivot, or files named .gdmp_pat, .gdmp_patg and .gdmp_sweep.

Wiz reports that an October 19 variant targeted self-hosted runners, including host files, process environments and cloud instance credentials. Some collection routines were broken. Their presence still warrants investigation; a buggy collector is no assurance that other theft failed.

Revoke access and replace exposed credentials

For affected repositories, Wiz recommends revoking the token that planted the workflow and rotating every accessible repository, organization and environment secret. Deleting the branch cannot revoke credentials already copied elsewhere.

If the malicious job executed on a self-hosted runner, treat the host and its instance role as compromised. Investigate through the established response process, including the token owner’s device and subsequent cloud activity.

For prevention, require reviewers for environments holding production secrets. Restrict OIDC, the identity exchange used for cloud access, to approved branches or environments. Prefer narrowly scoped, short-lived credentials and disposable runners with restricted host access.

Do not approve held sync-check runs or execute code from suspect branches. This is credential abuse, so the report supplies no product-version upgrade that resolves the exposure by itself.

Caveat: Wiz sells cloud-security products and supplies the campaign telemetry; its visibility does not establish the total size or success rate of Gheist.