What happened
Replaying the tj-actions/changed-files March 2025 intrusion
- 1
Upstream tokens leak
Dec 6, 2024 – Mar 11, 2025A pull_request_target workflow exposes an upstream maintainer's token on December 6, 2024, a second project's workflow run leaks a reviewdog maintainer's token on March 11, 2025 (an estimate), and the v1 tag of reviewdog/action-setup points to a malicious commit from 18:42 to 20:31 UTC.
- 2
The poisoned action
about Mar 10–14, 2025A token with write access to tj-actions/changed-files is taken, by a path reported three ways, and commit 0e58ed8 is added with most version tags repointed to it.
- 3
Secrets dumped into logs
Mar 12–15, 2025, per CISAWorkflows that reference the action by tag run the injected code, which reads the memory of the runner and prints the secrets to the log; a spear-targeted commit also reaches the Coinbase agentkit workflow on March 14.
- 4
Exposed cloud keys
Mar 14–15, 2025, per GitHubThe secrets of public repositories sit in public logs, valid AWS access keys among them; no use of any exposed key is reported.
Workflows that ran the action printed the secrets in the memory of their runner to the build log as double-encoded base64, and the logs of public repositories are public. Wiz identified dozens of affected repositories, some of large enterprises, and saw valid AWS access keys, GitHub personal access tokens, npm tokens and private RSA keys exposed; it observed no exfiltration to an attacker-controlled server, and no misuse of the exposed AWS keys is reported. The action is used in over 23,000 repositories, a partial count per Unit 42.
Where cloud-provider controls fit
If you run workflows on AWS, a service control policy that denies long-lived access key creation keeps new stored keys out of the memory dump, though a key created earlier stays valid, and OpenID Connect gives a workflow a short-lived credential in place of a stored one. The poisoned tag, the stolen token and the log belong to GitHub, where a ruleset on tags, an Actions policy that pins actions to a full commit SHA, environment secrets and read-only workflow tokens act. Wiz observed no exfiltration to an attacker server, and no source reports a use of an exposed AWS key.
Could this have been prevented?
In hindsight, not as a whole. 1 control would have refused a step outright, and 11 controls would have limited one. 8 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.
- 1
- Prevented
- 11
- Contained
- 8
- 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 tj-actions/changed-files March 2025 intrusion
1 · Upstream tokens leak
3 Contained3 Not covered2 · The poisoned action
2 Contained2 Not covered3 · Secrets dumped into logs
1 Prevented2 Contained2 Not covered4 · Exposed cloud keys
4 Contained1 Not covered3 on other cloudsKey facts
- Commit 0e58ed8, which showed a renovate bot as author but was unverified, was created outside the action's repository, and most version tags were repointed to it, so existing workflow references ran it; Unit 42 dates the push to about March 10 to 14, 2025
- On Linux runners the injected code downloaded a Python script from a GitHub Gist, ran it with sudo, read the memory of the Runner.Worker process and echoed the secret-shaped entries, encoded twice in base64, to the build log
- CISA dates affected use of the action from 00:00 UTC on March 12 to 12:00 UTC on March 15, 2025 and GitHub's advisory gives March 14 to 15; the action is used in over 23,000 repositories, a partial count per Unit 42
- Wiz saw valid AWS access keys, GitHub personal access tokens, npm tokens and private RSA keys exposed in what it calls dozens of repositories and observed no exfiltration to an attacker-controlled server; no misuse of the exposed AWS keys is reported
- Unit 42 reports that the Coinbase agentkit workflow ran the malicious code with write-all permissions at about 15:10 UTC on March 14 and leaked a token with write permissions; Coinbase reported the attack was unsuccessful at causing any damage
- StepSecurity flagged the compromise on March 14, GitHub removed the action on March 15 and restored it with all versions clean, and on March 26 CISA added CVE-2025-30066 and CVE-2025-30154 to its Known Exploited Vulnerabilities Catalog
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.Harden-Runner detection: tj-actions/changed-files action is compromisedStepSecurity · Mar 14, 2025 · ResearchDetection, malicious commit, tags repointed, payload behavior, double base64 in logs, response timeline
- 2.GitHub Action tj-actions/changed-files supply chain attack: everything you need to knowWiz Research · Mar 15, 2025 · ResearchSecret types seen in public logs, no exfiltration observed, the maintainers' account of the token, reviewdog link
- 3.tj-actions/changed-files supply chain attack (CVE-2025-30066)GitHub Advisory Database · Mar 15, 2025 · AdvisoryCVE, affected and patched versions, exposure window, payload description
- 4.Supply chain compromise of third-party tj-actions/changed-files (CVE-2025-30066) and reviewdog/action-setup (CVE-2025-30154)CISA · Mar 18, 2025 · AdvisoryCVEs, affected time windows, secret types, Known Exploited Vulnerabilities listing
- 5.Analysis of the tj-actions/changed-files and reviewdog compromisePalo Alto Networks Unit 42 · Mar 20, 2025 · ResearchUpstream chain, reviewdog compromise, token leaks, Coinbase target, payload variants, exfiltration status
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.