By Hamza Faisal
Cloud security company Wiz has revealed an automated campaign that turned stolen GitHub tokens into a secret-stealing machine, planting malicious workflows in more than 13,500 branches across over 850 compromised accounts. Wiz has tracked the campaign, which it calls Gheist after the “gheist” user agent the attackers’ tool sends with GitHub API requests, since September 15, 2026. The affected accounts range from individual developers’ personal projects to the private repositories of enterprise organisations, and the campaign was still active as of October 24.
What makes Gheist alarming is that it needed no new software vulnerability. It abused permissions that compromised accounts already held, and GitHub’s own automation system, to reach the credentials stored alongside code.
How the theft worked
Each attack started with a stolen GitHub access token, a credential that lets software act as a user. Using that token, the attackers created a branch named “ci-sync-” followed by six random characters, and added a single workflow file, .github/workflows/sync-check.yml, with the commit message “ci: sync config.” Pushing the branch triggered GitHub Actions, which ran the workflow and printed the job’s environment variables and every secret the repository could access.
The trick was in the encoding. GitHub normally masks recognised secret values in workflow logs, but the workflow encoded the secrets in base32, which defeated that verbatim filter. The attackers then downloaded the logs through GitHub’s own API and deleted the workflow runs within seconds. In enterprise organisations, Wiz says, the tool cycled through several tokens within minutes and cloned private repositories before planting the workflow.
The safeguard that stopped only half the attack
GitHub has a safeguard, introduced on July 28, 2026, that holds suspicious workflow runs for approval. Wiz says every public-repository run it checked was stopped by that control, held in an “action_required” state. The problem is that the control does not apply to private repositories or to GitHub Enterprise Server, and that is where the campaign succeeded.
In private repositories and on Enterprise Server, GitHub audit logs show runs on the planted branches finishing and then being deleted by the gheist user agent. AWS CloudTrail logs show the organisations’ own workflows also running on those branches and assuming AWS roles without approval, and hours later cloud credentials from the same organisation were used to enumerate AWS Secrets Manager. Wiz has not said how many organisations had secrets exposed or used, nor how the attackers obtained the tokens.
A second variant, added on October 19, targeted self-hosted runners, searching for 31 token formats — including Anthropic, Hugging Face, Groq and NVIDIA keys — alongside Kubernetes tokens and host filesystems reached through the Docker socket. Wiz says code bugs prevented Google Cloud and Azure token collection in that variant.
Still active — and what to do now
As of October 24, the campaign was still active and evolving, although roughly two-thirds of the identified accounts were no longer available on GitHub following Wiz’s report. Wiz’s advice is blunt: treat every secret that repository could read as exposed. Rotate those secrets, revoke the token that pushed the branch, and investigate affected devices and self-hosted hosts. Deleting the branch, or the run it started, does not undo the exposure, because the secrets were already printed into logs the attackers could download.
The broader lesson is uncomfortable. Gheist needed no external command server for firewalls to block: the platform’s own plumbing carried the secrets out. Where a single token can reach dozens of projects, keeping tokens short-lived, narrowly scoped and quickly revoked is no longer housekeeping — it is the security boundary itself.



