More screenshots
Works with the AI Agents you already use
GitHub Actions Security Audit: Find risky workflows
Maybe it adds `pull_request_target`, a write scope, or a step that echoes a secret.
$12
GitHub Actions Security Audit: Find risky workflows
Example session with this skill installed
Audit references/review.yml as public, report Markdown at the default high threshold, and prioritize fixes.
- Read your context and instructions
- Compiled the github actions security
Captured stdout from python3 -I -B scripts/gha_audit.py references/review.yml --visibility public --format markdown (exit 1):
GitHub Actions security audit
Files: 1 | Findings: 12 | Gate: FAIL
| File:line | Rule | Severity | Finding | Fix |
|---|---|---|---|---|
| references/review.yml:3 | excessive-permissions | high | All token scopes are writable. | Use an explicit minimal permissions mapping. |
| references/review.yml:6 | self-hosted-runner | high | Self-hosted runner on public or unknown-visibility repository. | Use hosted runners for untrusted jobs; verify repository visibility and isolation. |
| references/review.yml:7 | unpinned-image | medium | Container image lacks a SHA256 digest. | Pin the trusted image to @sha256:<64 hex digest>. |
| references/review.yml:9 | persist-credentials | medium | Checkout may retain token credentials for later steps. | Set with: persist-credentials: false unless later authenticated Git is required. |
| references/review.yml:9 | unpinned-action | high | Remote action or workflow lacks a full commit SHA. | Resolve and verify a trusted upstream commit; pin all 40 hex characters. |
| references/review.yml:10 | dangerous-checkout | critical | pull_request_target checks out attacker-controlled PR head. | Do not check out or execute PR head in this privileged event; use pull_request. |
| references/review.yml:13 | security-step-continue-on-error | medium | Security-named step may ignore failure. | Remove continue-on-error from required security gates. |
| references/review.yml:15 | script-injection | high | Event expression interpolated into a script. | Pass the value via env and use quoted environment variables; validate its contents. |
| references/review.yml:16 | secrets-echo | high | Secret expression sent to a printing command. | Remove secret logging; masking is not a safe logging policy. |
| references/review.yml:17 | deprecated-command | low | Deprecated workflow output/state command. | Use GITHUB_OUTPUT or GITHUB_STATE environment files. |
| references/review.yml:19 | unpinned-action | high | Remote action or workflow lacks a full commit SHA. | Resolve and verify a trusted upstream commit; pin all 40 hex characters. |
| references/review.yml:20 | secrets-inherit | high | Reusable workflow receives inherited secrets. | Pass only explicitly named secrets required by the called workflow. |
The gate fails with 12 findings. Address the privileged PR checkout first, then the secret printing and event interpolation. The included partial patch removes the latter two matches and leaves 10 findings, so the result still fails.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem
Someone opens a PR that touches .github/workflows/. Maybe it adds pull_request_target, a write scope, or a step that echoes a secret. You want a second pair of eyes before you merge, and you'd rather not install another tool just to get one.
See it before you buy: real input, real output Command (from the package folder, exit code 1, 0.025 s):
python3 -I -B scripts/gha_audit.py references/review.yml --visibility public --format markdown
Before: references/review.yml (excerpt)
on: pull_request_target
permissions: write-all
- uses: actions/checkout@v4
ref: ${{ github.event.pull_request.head.sha }}
echo "${{ github.event.pull_request.title }}"
echo "${{ secrets.EXAMPLE_TOKEN }}"
After: lines copied from the script's output (excerpt):
Files: 1 | Findings: 12 | Gate: FAIL
| references/review.yml:3 | excessive-permissions | high | All token scopes are writable. | Use an explicit minimal permissions mapping. |
| references/review.yml:9 | unpinned-action | high | Remote action or workflow lacks a full commit SHA. | Resolve and verify a trusted upstream commit; pin all 40 hex characters. |
| references/review.yml:10 | dangerous-checkout | critical | pull_request_target checks out attacker-controlled PR head. | Do not check out or execute PR head in this privileged event; use pull_request. |
| references/review.yml:15 | script-injection | high | Event expression interpolated into a script. | Pass the value via env and use quoted environment variables; validate its contents. |
| references/review.yml:16 | secrets-echo | high | Secret expression sent to a printing command. | Remove secret logging; masking is not a safe logging policy. |
What the numbers show
- Ran on 98 public workflow files from 18 repos: 51 were checked (the other 47 used YAML the reader doesn't support and were refused with exit 2 rather than guessed), with 148 findings.
- I read 20 findings by hand: 15 were worth a reviewer's attention and 5 were false alarms, all from one rule that flagged PR numbers and commit SHAs pasted into scripts.
- All 98 runs took about 3.1 seconds together, Python startup included.
What you get
- A single-file Python scanner (standard library only) with 15 checks: token permissions, action and container pins, PR-head checkout in privileged events, event text pasted into shell scripts, printed or inherited secrets, untrusted workflow_run artifacts, self-hosted runners on public repos, and security steps set to ignore failure.
- A JSON or Markdown report. Each finding has the file, line, severity, and a one-line fix.
- You can point it at a repo, a workflows folder, or one file, tell it whether the repo is public, and pick the gate level (low, medium, high or critical).
- Instructions for your agent: read the surrounding workflow, write a small unified diff with a reason for each change, never make up action SHAs, and rerun the same check after edits you approve.
- A worked patch on the demo file that takes it from 12 findings to 10 and leaves the rest visible.
- 21 command-line test cases and 4 saved sample reports.
Measured on real data (full detail)
I ran it on every .yml workflow in 18 public repositories (fastapi, axios, black, httpx and others, at pinned commits), one file at a time:
| What I ran | What happened |
|---|---|
| 98 public workflow files | 51 were read and checked. 47 used YAML the reader doesn't support, so it refused them with exit code 2 instead of guessing |
| The 51 checked files, default gate (fail on high) | 22 passed, 29 failed, 148 findings in total |
| All 98 runs back to back | about 3.1 seconds, Python startup included |
| 20 findings I read by hand | 15 were worth a reviewer's attention, 5 were false alarms |
Most findings were mutable action tags (60), write scopes (31), checkout keeping its token (27) and jobs with no permissions block (22). All 5 false alarms came from the same rule: it flagged PR numbers and commit SHAs pasted into scripts, which can't carry shell code. The 20 were hand-picked rather than random, and "worth attention" isn't the same as "exploitable". Several writes were clearly needed for releases. The pinned commits and every label are in the benchmark notes inside the package.
Known limitations
- The YAML reader is deliberately small. It refused 47 of the 98 public files, and those got no check at all.
- It doesn't follow
if:conditions, matrices, reusable workflow contents, local actions or external scripts. - It checks that a pin looks like a full SHA. It can't tell you whether that SHA is the commit you meant.
- It flags every event field put into a script, so PR numbers and SHAs show up as false alarms.
- A PASS only means nothing reached your gate level. It isn't a security sign-off.
Who it's for Maintainers and reviewers who want a quick local check on workflow changes, plus a drafted patch, without adding a dependency. If you can install a dedicated Actions linter, use it alongside this or instead of it.
Requires Python 3.8 or newer and an agent that can run commands and read files. No Git, no GitHub account, no packages to install. Checking that a SHA matches a real release needs a trusted source you provide.
Privacy: everything runs on your machine. No network, no telemetry, and the scanner writes no files. Your agent only edits what you ask it to.
Price and guarantee $12 one time, yours to use with any agent that can run Python. Covered by Agensi's money-back guarantee if it doesn't do what this page promises.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
No shell parsing or runtime monitoring. Skips matrices and local actions. 1MB file size limit. Requires Python 3.8+. No network-based provenance verification for action SHAs.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 15 days ago