What happened
Replaying the Gheist September 2026 intrusion
- 1
Stolen tokens plant workflows
Tracked from Sep 15, 2026Stolen tokens of users and automation accounts push a new branch with a workflow file to each repository they reach; how the tokens were stolen is unconfirmed.
- 2
Secrets printed to the run log
Environment secrets from Sep 28, 2026The workflow prints the repository's secrets into its run log in an encoded form that log masking misses; the tool downloads the log through the GitHub API and deletes the run.
- 3
Workflows assume AWS roles
Not datedIn private and GitHub Enterprise repositories the organizations' own workflows run on the attacker branches and assume AWS roles; hours later, in one enterprise, the organization's cloud credentials enumerate Secrets Manager.
- 4
Self-hosted runner job
Added Oct 19, 2026A second job for repositories with a registered self-hosted runner targets AWS instance-role credentials and cloud metadata; whether it ran on any victim runner is unconfirmed.
Secrets of repositories reached by more than 850 compromised accounts were printed into run logs and taken through the GitHub API; the number of organizations with exposed secrets is not stated. In private and enterprise repositories, workflows on attacker branches assumed AWS roles, and in one enterprise the organization's cloud credentials enumerated Secrets Manager. Every public-repository run Wiz checked was held and did not run.
Where cloud-provider controls fit
On AWS, a role trust policy whose sub condition names one repository and one protected branch, with no wildcard, would have refused the role to a workflow on an attacker branch. A Microsoft Entra federated identity credential or a Google Cloud attribute condition answers the same way. An Oracle Cloud identity propagation trust maps only the tokens its rules match, and Oracle's documentation does not name GitHub tokens. A service control policy with network conditions would have refused the Secrets Manager listing from outside the organization's networks, if the account is a member account, and would have capped where a runner's instance-role credentials work. The rest sits with GitHub. An organization IP allow list would have refused a stolen token used from an address outside the list, if the organization uses GitHub Enterprise Cloud. A ruleset that restricts branch creation and a personal access token policy narrow what a stolen token can push. Environment secrets behind required reviewers, and OpenID Connect in place of stored cloud keys, keep those secrets out of a printed log. No cloud control reaches the token or the repository-level secrets, so you own which tokens can write to your repositories and which secrets a workflow on a new branch can read.
Could this have been prevented?
In hindsight, not as a whole. 3 controls would have refused a step outright, and 6 controls would have limited one. 2 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.
- 3
- Prevented
- 6
- Contained
- 2
- Not covered
How it could have been prevented
The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.
Replaying the Gheist September 2026 intrusion
1 · Stolen tokens plant workflows
1 Prevented2 Contained1 Not covered2 · Secrets printed to the run log
2 Contained1 Not covered3 · Workflows assume AWS roles
2 Prevented1 Contained6 on other clouds4 · Self-hosted runner job
1 Contained3 on other cloudsKey facts
- Wiz Research began tracking the campaign on September 15, 2026 and named it after the user agent of its GitHub API requests; it named no victim and attributed the campaign to no actor
- The tokens belonged to users and automation accounts; Wiz assessed that the ones it could trace fit infostealer logs and tokens exposed in earlier supply-chain compromises, and the token type was not stated
- For each repository a token could reach, the tool created a branch with a randomized name and committed an unsigned workflow file authored as the token's owner; the workflow printed the repository's secrets into its run log, encoded so that log masking missed them, and a tool version first seen on September 28 printed environment-scoped secrets too
- The tool downloaded each run log through the GitHub API and deleted the run seconds later; no attacker server or outbound connection was involved
- GitHub began holding potentially malicious workflow runs in public repositories for approval on July 28, 2026, and every public-repository Gheist run Wiz checked stayed held; in private and GitHub Enterprise repositories runs on attacker branches completed, and CloudTrail showed the organizations' own workflows on those branches assuming AWS roles
- By October 24, 2026 Wiz had recovered more than 13,500 planted branches and counted more than 850 compromised accounts, and the campaign was ongoing; in one enterprise the organization's cloud credentials enumerated AWS Secrets Manager hours after the workflows ran
Sources
Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.
- 1.Gheist: Attempted Scaled Abuse of Stolen GitHub Tokens StumblesWiz Research · Oct 25, 2026 · ResearchTimeline, token use, workflow mechanics, secret exfiltration, tool versions, AWS role assumptions, Secrets Manager enumeration, self-hosted runner job, counts
- 2.GitHub Actions holds potentially malicious workflows for approvalGitHub · Jul 28, 2026 · Vendor reportApproval hold, scope of the hold
- 3.Gheist turns GitHub Actions into a secret-extraction serviceCSO Online · Oct 28, 2026 · NewsCorroboration of counts, Secrets Manager enumeration, CloudTrail role assumptions
The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.