ciguard

Security auditor for GitHub Actions workflows. Point it at a workflow file or a repo and get a risk report: script injection, unpinned third-party actions, pull_request_target footguns, and overly broad permissions.

Why

CI pipelines are production infrastructure with the keys to the kingdom, and most workflow files were written by copying Stack Overflow. Supply-chain attacks through CI are some of the highest-leverage attacks out there, and reviewing every workflow by hand doesn't scale. ciguard automates the first pass so security teams can focus on the actually suspicious stuff.

Install

npm install -g ciguard

Or run from source:

git clone https://github.com/burrejak22/ciguard
cd ciguard
npm install

Usage

# single workflow file
ciguard .github/workflows/ci.yml

# whole repo (walks for *.yml / *.yaml)
ciguard ./my-repo

# machine-readable output
ciguard ./my-repo --json

CI usage

Fail the build when risk hits a threshold:

ciguard .github/workflows --fail-on HIGH

Exit code is 1 when the worst finding across all scanned files is the given band or worse (LOW < MEDIUM < HIGH < CRITICAL).

Example output

ci.yml
[HIGH] risk score: 60/100 ████████████░░░░░░░░

  ▸ +40  TRIGGER_PR_TARGET
     pull_request_target runs in a privileged context and checks out + runs PR code — classic pwn-request setup
  ▸ +35  SCRIPT_INJECTION
     script injection: attacker-controlled context interpolated into run: (job build, step 2)
  ▸ +20  ACTION_UNPINNED
     some-random/action@v1 uses a mutable tag — pin to a sha for supply-chain safety (job build)

What gets checked

• Script injection — ${{ }} expressions inside run: scripts, weighted by whether the context is attacker-controlled (github.event.issue.title, github.head_ref, ...). Worse under pull_request_target.
• Unpinned actions — uses: refs on mutable tags or floating branches instead of a commit sha. Third-party actions score higher than actions/*.
• Dangerous triggers — pull_request_target combined with checking out and running PR code (the pwn-request pattern).
• Permissions — write-all at the workflow or job level.
• Secret exfiltration — secrets echoed into logs or shipped to external URLs via curl.
• Cache poisoning — actions/cache keys built from attacker-controlled data on PR triggers.
• Credential persistence — actions/checkout keeping the token around while untrusted PR code runs.

Scoring

Same philosophy as the rest of the family: weights are opinionated triage heuristics, not a certification. Totals capped at 100.

| Band | Score | Meaning | |----------|--------|-----------------------------------| | LOW | 0–19 | boring in a good way | | MEDIUM | 20–44 | worth a look | | HIGH | 45–69 | needs justification | | CRITICAL | 70–100 | fix before this runs again |

Limitations

• Static analysis only. It reads the YAML, it doesn't execute your pipeline.
• Expression analysis is heuristic — complex nested contexts may be missed, and safe patterns can occasionally get flagged. A HIGH score means "a human should look", not "you've been pwned".

Contributing

Issues and PRs welcome. Keep it dependency-light and keep the CLI fast.