In March 2025 an attacker moved the version tags of tj-actions/changed-files, a GitHub Action used by more than 23,000 repositories, to a malicious commit. Every workflow that referenced the action by tag, from @v1 to @v45.0.7, then ran code that read the runner's memory and printed the secrets it found into the job log. It is tracked as CVE-2025-30066, and CISA added it to its Known Exploited Vulnerabilities catalog on 18 March 2025. The way in was another action, reviewdog/action-setup (CVE-2025-30154), compromised three days earlier.
It is the incident that made "pin your actions to a commit SHA" standard advice. Nothing in the affected workflows changed: the same uses: line that was safe on Thursday ran malicious code on Friday. And the attacker didn't need a server of their own. In a public repository, anyone could read the log. A year later, the Trivy supply chain attack used the same technique against trivy-action.
What happened
According to Unit 42, Wiz, the advisories and CISA:
| When (UTC) | What happened |
|---|---|
| 28 Nov to 6 Dec 2024 | A SpotBugs maintainer added a personal access token to a SpotBugs workflow. The attacker leaked it with a malicious pull request against a workflow that used the pull_request_target trigger. |
| 11 March 2025 | With that token, the attacker pushed a malicious workflow to SpotBugs that leaked the token of a reviewdog maintainer, and moved reviewdog/action-setup@v1 to a malicious commit between 18:42 and 20:31. |
| 11 March 2025 | tj-actions/eslint-changed-files used reviewdog/action-setup@v1. When it ran in tj-actions' CI, it leaked a token with write access to tj-actions/changed-files, which Wiz calls the likely root cause. |
| 12 to 14 March 2025 | Unit 42 says the attacker first aimed at Coinbase: it forked coinbase/agentkit, and at 15:10 on 14 March a Coinbase workflow ran a malicious changed-files commit and leaked a token with write permissions. Coinbase removed the workflow; Unit 42 found no damage to any Coinbase asset. |
| 14 March 2025, 16:57 | The attacker moved the changed-files tags v1 to v45.0.7 to the malicious commit 0e58ed8. |
| 14 March 2025 | StepSecurity reported it, after its monitoring flagged an unexpected connection to gist.githubusercontent.com. |
| 15 March 2025 | GitHub published GHSA-mrrh-fwg8-r2c3 and CVE-2025-30066. The fix shipped in 46.0.1, and the tags were restored to safe commits. |
| 18 to 24 March 2025 | CISA published an alert (18 March) and added CVE-2025-30066 (18 March) and CVE-2025-30154 (24 March) to the KEV catalog. reviewdog published its announcement and advisory. |
Who is affected
tj-actions/changed-files: every workflow run that used the action by tag while the tags pointed to0e58ed8, which the CVE record dates to 14 and 15 March 2025, and every run that referenced0e58ed8directly. CISA's window for auditing is wider: 12 March 00:00 to 15 March 12:00 UTC.reviewdog/action-setup@v1: every run between 18:42 and 20:31 UTC on 11 March 2025. The advisory adds thatreviewdog/action-shellcheck,action-composite-template,action-staticcheck,action-ast-grepandaction-typoswere compromised in that window "regardless of version or pinning method", because they callaction-setup@v1themselves.- Who was exposed most: public repositories, where anyone could read the workflow logs. In private repositories, the secrets reached everyone with read access to the logs and any system that collects them.
Workflows pinned to a full commit SHA of changed-files other than 0e58ed8 were not affected by the tag move.
How it works
Both CVEs are classified as embedded malicious code (CWE-506). The mechanism:
- Tags are mutable. A Git tag is a name pointing at a commit, and whoever has write access can point it elsewhere. A workflow that says
uses: tj-actions/changed-files@v44resolves the tag every time it runs. GitHub's secure use reference recommends a workflow lockfile, or a full-length commit SHA, so a run uses exactly the commit you reviewed. - The payload read the runner's memory. According to GitHub's advisory, the malicious commit downloaded a Python script from a GitHub gist and ran it with
sudo. The script read the memory of theRunner.Workerprocess, which holds every secret the job has, including secrets never passed to that step. - It printed the secrets into the log. The output was base64-encoded twice, so it no longer matched the secret values GitHub masks in logs, and printed in the action's output. No exfiltration server was needed; the log was the channel.
- reviewdog hid it in the installer. In
reviewdog/action-setup, Wiz found the payload base64-encoded and inserted directly into theinstall.shthe action runs. - Nested actions inherit the risk.
eslint-changed-filesand the other reviewdog actions calledaction-setup@v1by tag. Pinning the outer action did nothing for the inner one.
How to check whether you are affected
The workflow logs of March 2025 are gone: GitHub deletes them after 90 days by default and after 400 days at most. So the question is what your workflows referenced at the time, and which runs used them.
- Find the references as they were in March 2025, on every branch that ran CI then, not as they are today:
tj-actions/changed-files,tj-actions/eslint-changed-files,reviewdog/action-setupand the reviewdog actions listed above, including inside your own composite actions and reusable workflows. - Sort them by pin. A tag reference ran the malicious commit if it ran in the window. A SHA reference was safe unless it was
0e58ed8. Any version of the five reviewdog actions was affected in reviewdog's window. - List the runs in the window. The run history shows which workflows ran between 11 and 15 March 2025, even where the logs are gone.
- If you have logs, from an archive or a log platform, search the
changed-filesstep for long base64 strings and decode them withbase64 -d | base64 -d, as the advisory suggests. A secret in the output means it was exposed to everyone who could read that log.
Commands you can adapt. To see the workflow files as they were on 15 March 2025 on every remote branch:
git fetch --all --quiet
for ref in $(git for-each-ref --format='%(refname:short)' refs/remotes); do
commit=$(git rev-list -1 --before='2025-03-15T12:00:00Z' "$ref") || continue
[ -n "$commit" ] && git grep -n -E 'tj-actions/(changed-files|eslint-changed-files)|reviewdog/action-' "$commit" \
-- '.github/workflows/*' '.github/actions/*' '*action.yml' '*action.yaml' | sed "s|^|$ref |"
done
To list the runs of a workflow in the window (replace ORG/REPO and the workflow file):
gh run list --repo ORG/REPO --workflow ci.yml --created '2025-03-11..2025-03-15' \
--limit 1000 --json databaseId,headBranch,event,createdAt \
--jq '.[] | [.databaseId, .headBranch, .event, .createdAt] | @tsv'
Why today's workflow file doesn't settle it
- The file never changed.
@v44reads the same before and after the attack. Only the commit it resolved to on the day tells you what ran. - Today's branch isn't the one that ran. A pull request branch from March 2025, since merged or deleted, ran CI with its own workflow files.
- Nested references hide. A composite action or reusable workflow of your own, or a third-party action, may have called
changed-filesoraction-setupwithout your workflow naming it. - The logs that would prove it are gone. Without them, a run that used an affected tag in the window has to be treated as exposed.
What you need is a record of which action commits each run actually used, kept per run. See are we affected? how to assess a new CVE across your builds.
How to fix it, and what to do if you can't patch yet
- Rotate the secrets of every affected run. CISA's advice: "Rotate all identified secrets immediately as they should be considered compromised." Rotate everything the job could read, since the payload read the runner's memory: cloud keys, personal access tokens, package publish tokens and private keys. In a public repository, assume the log was read.
- Update. Use
tj-actions/changed-files46.0.1 or later, andreviewdog/action-setupat or after commit3f401fe, as CISA advises. Remove any reference to0e58ed8. - Pin to full commit SHAs, for
changed-filesand every other third-party action, and check that the actions you use pin what they call.
Verify the fix
- No tag references remain for third-party actions: every
uses:line outside your own organization has a 40-character SHA. - No reference to
0e58ed8exists on any branch. - The old credentials no longer work. Revoked tokens and keys are deactivated at the provider, not just replaced in the repository's secrets.
- The next run is clean. Check the commit each action resolved to in the run, not only the workflow file.
How to reduce the damage next time
- Enforce SHA pinning. Since August 2025, GitHub's allowed actions policy can require full commit SHAs and block specific actions or versions. GitHub's workflow lockfiles, generally available since 28 October 2026, record the commits of actions and their transitive dependencies, the gap that reviewdog's nested actions fell through.
- Never run untrusted code with secrets. The chain started with
pull_request_targetrunning a fork's code with a maintainer's token. See GitHub Actions security and GitHub Actions secrets: how they leak and how to lock them down. - Give each job only the secrets and token permissions it needs. A job that lists changed files needs no cloud keys. See GITHUB_TOKEN permissions.
- Limit where builds can connect. The payload downloaded its memory-dumping script from
gist.githubusercontent.com, and the outbound call is what StepSecurity spotted. A job that can reach only the hosts it needs can't fetch it; see how an egress policy stops exfiltration from GitHub Actions. This doesn't help when the payload is inline, as in reviewdog'sinstall.sh, or when secrets leave through the log.
Check for compromise
From the advisories and CISA:
| Type | Indicator |
|---|---|
| Commits | tj-actions/changed-files 0e58ed8671d6b60d0890c21b07f8835ace038e67; reviewdog/action-setup f0d342d (malicious) and 3f401fe (fix) |
| Network | A download from gist.githubusercontent.com (memdump.py) during the changed-files step |
| Logs | Long double base64-encoded strings in the output of the changed-files or reviewdog steps, which decode to secrets |
When to assume compromise. Every run that used an affected reference in the window. GitHub's advisory asks you to rotate secrets "as a precaution" even where you found nothing, and CISA treats identified secrets as compromised.
Review what the credentials touched. For every exposed credential, check cloud, registry and GitHub audit logs for use you can't explain from March 2025 onwards: new tokens, pushes, releases and workflow changes.
What to tell your customers
If you publish software built with GitHub Actions, customers may ask whether this incident reached your release pipeline. Answer with what you checked and what you did:
We checked every workflow that ran between 11 and 15 March 2025 for tj-actions/changed-files and reviewdog actions affected by CVE-2025-30066 and CVE-2025-30154. [No workflow used an affected version.] / [Workflows in [repository] used an affected version on [date]. We rotated every credential those runs could access, reviewed audit logs for misuse and found [no evidence of misuse / the following: ...].] All third-party actions in our release pipeline are now pinned to full commit SHAs.
For templates for each outcome, see how to answer customers after a supply chain attack.
How CRACI helps
CRACI runs your GitHub Actions jobs on its own runners (runs-on: craci) and records what each job actually used: every package it fetched, and the GitHub Actions it ran, recorded as pkg:github components pinned to the commit used. A workflow line that says @v44 doesn't tell you what ran on the day; the job's record does, so a moved tag shows up as a different commit in the builds that ran it. Build history keeps every past build with its SBOM, so you can see which builds contained a component and when, within your retention period (180 days by default), and each job's network trace shows what it connected to.
CRACI's egress policy limits where a job can connect, is validated before the job starts and fails closed, so a payload fetching a script or sending secrets to a host the build doesn't need is blocked, even for an attack nobody has seen before. It doesn't stop secrets printed into the job's own log. Known malicious releases with a published CVE are flagged like any other vulnerability, and policy gates can block builds that contain them. CRACI doesn't detect malicious code by behavior or reputation, so a new compromise isn't flagged before an advisory exists. For the wider picture, see CI/CD security.
See the record of one of your own builds
Other incidents: all vulnerabilities, including the Trivy supply chain attack, the LiteLLM PyPI compromise that followed it, and the Shai-Hulud npm worm.
Sources and updates
Primary sources and original research:
- GitHub: GHSA-mrrh-fwg8-r2c3 (CVE-2025-30066), secure use reference
- reviewdog: GHSA-qmg3-hpqr-gqvc (CVE-2025-30154), maintainer announcement
- tj-actions: issue #2463
- CISA: alert of 18 March 2025, KEV catalog
- NVD: CVE-2025-30066, CVE-2025-30154
- Wiz on reviewdog/action-setup and Unit 42 on the full chain
- FIRST: EPSS scores
Updates:
- 2026-11-16: First published.