Gheist Turned GitHub Actions Logs Into a Secret-Exfiltration Channel
Gheist used stolen GitHub tokens to expose CI secrets through Actions logs, revealing a dangerous gap around private repositories and self-hosted runners.
A credential-theft campaign has been turning compromised GitHub access into a conveyor belt for CI/CD secrets. Wiz Research says the operation, which it calls Gheist, planted more than 13,500 branches across more than 850 compromised accounts between September 15 and October 24, 2026. The affected accounts ranged from personal projects to private enterprise repositories.
The scale matters, but the mechanism matters more. Gheist did not need a custom command-and-control server inside each victim's network. It used GitHub itself: a stolen token pushed a workflow, the workflow exposed secrets in an Actions log, and the attacker retrieved and quickly deleted that log through the GitHub API. Repository access became a path to cloud credentials, model-provider keys, signing material, and whatever else the build system could reach.
The numbers and victim observations come from Wiz and have not been independently reproduced. GitHub has not published a Gheist-specific incident report. GitHub has, however, documented the security control that Wiz says stopped the public-repository portion of the campaign, and the limitation that left private and self-hosted environments exposed.
A stolen token became a secret-extraction tool
According to Wiz, the automation created branches matching ci-sync- followed by six lowercase letters or digits. It then committed .github/workflows/sync-check.yml with the messageci: sync config. The workflow attempted to serialize available secrets, transform them with base32 encoding, and print the result to its own job log.
GitHub masks known secret values when they appear verbatim in logs. Encoding changes the representation, so literal-value masking is not a complete defense against a workflow that is already authorized to read a secret. Gheist's operator could download the output through the normal Actions API and delete the run seconds later. Wiz says the API client identified itself with the user agent gheist, giving defenders another useful indicator.
That sequence is a reminder that a CI log is an egress channel even when no unfamiliar domain appears in network telemetry. The attacker did not have to send secrets directly from a runner to outside infrastructure. The workflow stored them in a trusted service, then collected them through an authenticated control-plane request.
GitHub's public-repository hold worked
On July 28, GitHub announced that Actions would hold certain potentially malicious workflow runs in public repositories until a collaborator with write access approved them through an authenticated browser session. The protection is automatic. GitHub explicitly said it applies only to public repositories on github.com and does not apply to GitHub Enterprise Server.
Wiz reports that every Gheist run it examined in public repositories was left in anaction_required state. That is a meaningful defensive success: the attacker had valid credentials and could write a workflow, yet an execution-time control still interrupted the next step. It also shows why credential security alone is an incomplete model. A stolen token should be treated as probable initial access, not automatic permission to run arbitrary automation.
The boundary of that protection is just as important. Wiz says it observed successful runs in private repositories and GitHub Enterprise Server environments. In enterprise cases, credentials associated with the same organizations were later used to enumerate AWS Secrets Manager. That follow-on activity does not prove every exposed credential was used, but it shows the plausible blast radius extends beyond source code.
Self-hosted runners multiply the damage
The campaign evolved. Wiz found a later job aimed at self-hosted runners that searched beyond the GitHubsecrets context. It looked for environment variables, private keys, .env files, Terraform state, shell history, Kubernetes service-account tokens, runner credentials, cloud-instance metadata, and access to the Docker socket. It also included patterns for credentials from Anthropic, Hugging Face, Groq, Nvidia, and other services.
Some collection routines contained implementation mistakes and did not work as intended, according to the researchers. That limits what can be inferred from the code alone. A failed collector is not evidence that a runner was safe, though. If an unauthorized workflow executed on a persistent self-hosted machine, the host and every identity available to it must be investigated as potentially compromised.
This is the architectural danger of a durable runner. Build systems accumulate context: cached packages, deployment keys, cloud roles, source checkouts, local configuration, histories, and sockets into other privileged systems. That continuity is convenient for performance, but it turns a single malicious job into a search across yesterday's operational residue.
The incident is a provenance problem
Security teams often review what a workflow does while giving less attention to how it arrived. Gheist exploited that gap. The malicious file appeared in the repository through credentials belonging to a real user or automation identity. The relevant evidence is therefore not only the YAML body; it is the full history around it: token identity, branch creation, commit message, workflow state, runner assignment, log access, deletion, and subsequent cloud activity.
The same principle applies to AI agents that can edit repositories or trigger automation. A system that remembers prior instructions or credentials must also preserve their origin, scope, and expiration. Memory without provenance can turn an old authorization into a current capability. The appropriate response is not to remove persistence, but to bind remembered authority to auditable constraints and require fresh approval at high-impact boundaries.
What defenders should do now
Organizations should hunt for the branch pattern ^ci-sync-[0-9a-z]{6}$, the commit message ci: sync config, the path .github/workflows/sync-check.yml, and API requests using the gheist user agent. Because runs and branches may have been deleted, audit logs and local mirrors matter more than the current repository view.
Finding a match calls for more than deleting the workflow. Revoke the compromised GitHub token, rotate every secret the job could access, and review cloud control-plane activity for follow-on enumeration or use. If a self-hosted runner executed the workflow, isolate and rebuild it, examine the instance role and Kubernetes identity, and treat exposed signing or deployment credentials as compromised.
The longer-term controls are structural: short-lived GitHub App or OIDC credentials instead of broad personal tokens; environment protection rules with required reviewers; OIDC trust restricted to protected branches and environments; ephemeral self-hosted runners; and no unrestricted Docker socket or instance metadata access. Workflow creation and modification should generate high-signal alerts, especially when a new branch immediately introduces automation.
Gheist is not evidence that GitHub Actions is uniquely unsafe. It is evidence that CI/CD systems are privileged execution environments whose logs, identities, and retained state deserve the same controls as production infrastructure. GitHub's public approval hold broke the attack chain where it applied. Private repositories and self-hosted runners now need an equivalent assumption: a valid credential can be hostile, and execution should still have to earn trust.
Sources
Your AI remembers everything. Everywhere.
Unimatrix gives you a shared, durable memory layer across Claude Desktop, Cursor, ChatGPT, and Gemini. Setup in 2 minutes. Free and paid plans available.
Looking for developer resources?
Browse our catalog of 500+ tested AI prompt profiles covering DevOps, data modeling, agent behaviors, and API wrappers. Have a prompt to share? Submit your own for review by our librarian to be featured. Completely free, no registration required. Browse prompt libraries →
Keep reading
CERT-In says CRAFTY FALCON is re-attacking financial firms with agentic tools, exposing weak remediation, cloud controls, and trusted payment APIs in India.
Autoware 2.0 integrates learned driving planners, but its proposed runtime safety selector does not yet operate in Autoware’s present open-source release.
A Nature paper demonstrates a designed molecular spin–photon interface, but scalable quantum chips still require integration, entanglement, and lower loss.