What happened
Replaying the GhostAction September 2025 campaign
- 1
Account takeover
Sep 2, 2025Compromised maintainer accounts push a workflow named Github Actions Security to the repositories; how the accounts were compromised is not stated.
- 2
Workflow reads the secrets
Sep 2 to 5, 2025The workflow runs on every push and on manual dispatch and reads the repository's Actions secrets by the names the attackers collected.
- 3
Secrets leave
Sep 2 to 5, 2025The workflow posts the secrets by HTTP POST to a primary attacker-controlled host, with a second domain used briefly; 3,325 secrets were taken, an estimate.
- 4
Stolen cloud keys used
Not stated, reported by Sep 5, 2025Affected developers confirm that attackers use stolen AWS access keys and database credentials; the sources give no count, account, time or action.
Malicious workflows read and posted CI secrets from 817 repositories of 327 GitHub users, 3,325 secrets in all, an estimate, most often DockerHub credentials, GitHub tokens and npm tokens, per GitGuardian. GitGuardian says affected developers confirmed attackers using stolen AWS access keys and database credentials, and gives no count, account or time. No malicious package release was reported, and 9 npm and 15 PyPI packages were at risk.
Where cloud-provider controls fit
On AWS, a service control policy that refuses new long-lived access keys, with OpenID Connect from GitHub Actions in their place, would have kept new AWS keys out of the repository's secrets for the workflow to read, and a trust policy conditioned on the repository and environment would have capped the role it assumes; Google Cloud workload identity federation follows the same pattern. Where a key still leaked, a source-network condition would have refused it from the attacker's network, and an egress allow list would have refused the POST only from a self-hosted runner in your own network; the sources do not say which runners ran the workflow. The rest sits with GitHub and the registries: a ruleset that holds a new workflow to a reviewed pull request unless the account can bypass it, environment secrets behind required reviewers, npm trusted publishing and the npm setting that disallows tokens. The DockerHub, PyPI and database secrets stay in the secret store, and no cloud control reaches the egress of a GitHub-hosted runner, so you own which secrets a repository holds and who may push a workflow.
Could this have been prevented?
In hindsight, no control would have stopped it. 10 controls would have limited how far it went. 4 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.
- 0
- Prevented
- 10
- Contained
- 4
- 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 GhostAction September 2025 campaign
1 · Account takeover
1 Contained2 Not covered2 · Workflow reads the secrets
6 Contained1 Not covered3 on other clouds3 · Secrets leave
1 Contained1 Not covered3 on other clouds4 · Stolen cloud keys used
2 Contained3 on other cloudsKey facts
- A compromised maintainer account of the FastUUID project pushed the first malicious commit on September 2, 2025 and the attacker was then inactive for three days; identical commits went to at least 5 other public repositories and an estimated 10 private ones the same day, per GitGuardian
- The workflow .github/workflows/github_actions_security.yml ran on push and on manual dispatch, read the secrets by names the attackers had listed from each repository's own workflows and sent them by HTTP POST to an attacker-controlled host on a plesk.page subdomain; all affected repositories used the same primary endpoint and a second exfiltration domain was used briefly
- 327 GitHub users and 817 repositories were affected and 3,325 secrets were taken, an estimate that BleepingComputer calls rough; the most common were DockerHub credentials, GitHub tokens and npm tokens, and the others included PyPI tokens, AWS access keys, database credentials and Cloudflare API tokens
- GitGuardian says initial discussions with affected developers confirmed active exploitation of stolen AWS access keys and database credentials; it gives no key count, no account, no time and no action
- 9 npm and 15 PyPI packages were at risk of compromise, per GitGuardian, and the FastUUID PyPI token was stolen, per SecurityWeek and BleepingComputer, but no malicious release was reported; PyPI set the FastUUID project to read-only on September 5
- On September 5 GitGuardian notified the GitHub, npm and PyPI security teams and opened issues in 573 of the 717 repositories that had not reverted the change, and the exfiltration hostname stopped resolving at 16:15 UTC, per StepSecurity; about 100 repositories had already reverted
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.The GhostAction Campaign: 3,325 Secrets Stolen Through Compromised GitHub WorkflowsGitGuardian · Sep 5, 2025 · ResearchTimeline, counts, injection method, exfiltration behavior, secret types, notifications, AWS key abuse, unconfirmed points
- 2.GhostAction Campaign: Over 3,000 Secrets Stolen Through Malicious GitHub WorkflowsStepSecurity · Sep 5, 2025 · ResearchWorkflow mechanics, exfiltration endpoint, indicators
- 3.GitHub Workflows Attack Affects Hundreds of Repos, Thousands of SecretsSecurityWeek · Sep 8, 2025 · NewsDiscovery date as reported, FastUUID token, secret types as reported
- 4.Hackers steal 3,325 secrets in GhostAction GitHub supply chain attackBleepingComputer · Sep 8, 2025 · NewsTimeline as reported, counts as estimates, unconfirmed points
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.