How to check if an agent skill is safe before installing

An AI-agent skill runs on your machine with your agent's permissions, so the time to look at it is before it is installed. This page is the procedure, with SkillScan by RakTikoo as the worked example: install the scanner once, scan the skill folder, and read one plain-English verdict. The scan is offline and deterministic — nothing leaves your machine, the scanned code is never executed, and the same folder always gets the same answer.

1. Install the scanner

pip install git+https://github.com/RakTikoo/skillscan

(Python 3.10+, zero dependencies. Please use the GitHub route: the skillscan name on PyPI belongs to a different tool by another author. Other routes — pipx, uv, a release wheel, Homebrew, Scoop — are on the index page.)

2. Scan the skill folder

Point the scanner at the folder you are thinking of installing:

skillscan <path-to-skill-folder>

You can also scan a GitHub repo directly — SkillScan shallow-clones it into a temp dir, scans it, and deletes the clone:

skillscan https://github.com/<owner>/<repo>

SkillScan walks the folder and scans every regular file that decodes as text, whatever its extension. Every report states how many files were scanned and which were skipped; if anything suspicious was skipped — or nothing was scanned at all — the verdict cannot be a clean PASS.

3. Read the verdict

A scan of a clean skill ends like this:

No known-bad patterns found.

VERDICT: PASS

A scan of a skill that reads your SSH keys and sends data out over the network ends with findings that name the rule, the file, and the line — for example SSH_CREDENTIAL_READ and EXFIL_COMBINATION — and:

VERDICT: BLOCK

The three verdicts, and what to do with each:

VerdictExit codeMeaning
PASS0No known-bad patterns were found in what was scanned. That is not a guarantee the skill is safe — see the limits below.
CAUTION1Something worth a human read before installing. Read the findings and decide whether each one makes sense for what the skill claims to do.
BLOCK2At least one block-severity pattern was found. Do not install until you have read the findings and judged the context for yourself.

Exit code 3 means the scanner itself could not run (bad path, bad flags) — deliberately distinct from BLOCK. Because the verdicts are exit codes, the same scan works as a CI / install-time gate.

Reports also come in other formats: skillscan <folder> --format md for Markdown, --format json for machine-readable output, and --format sarif for SARIF 2.1.0 toolchains.

Honest limits

SkillScan is a linter, not a guarantee of safety. A PASS means "no known-bad patterns were found in the files that were scanned" — not "this skill is safe". Runtime behaviour is invisible to a static scan, and deeply dynamic code can slip through. The full limits list is on the index page and in the repository. SkillScan's own measured detection and false-block rates — including the families where the rates are poor — are on the measured results page.

Already installed a skill and want to know whether it has changed since? See Did this skill change since I installed it?