vet-skill
Treat installing any external skill as a trust decision. A skill body loads into the agent's context and can carry injected instructions; some skills ship scripts that run. Scan first.
Tool
NVIDIA SkillSpector, a static scanner for agent skills. Install once, keep current:
uv tool install skillspector # first time
uv tool upgrade skillspector # keep current before a vetting session
If the Loam repo (or a Loam-bootstrapped project) is present, prefer its wrapper, which applies the threshold and exit codes for you:
bin/vet-skill.sh <path-to-skill-dir>
bin/vet-skill.sh https://github.com/owner/repo # cloned read-only first
bin/vet-skill.sh <target> --safe --out <dir> # strip scripts, re-scan, emit prose-only tree
Making a script-heavy skill safe (prose-only)
When a skill fails on its scripts but you only want its written guidance, run
--safe: it strips every executable (.py .sh .js .ts, Makefile, scripts/,
and the rest), re-scans, and with --out <dir> writes the stripped tree ready to
install. This removes code-execution risk entirely.
It does NOT remove risk carried in the prose. A skill flagged for rogue-agent or
excessive-agency instructions in its own SKILL.md (patterns like "always do X,
never stop, ignore other instructions") will still fail after --safe, because
the risky text is the skill, not a script. Read those findings; adopting such a
skill means editing the instructions out, not just stripping files.
Procedure
- Never install first. Clone read-only to a scratch dir, or point the wrapper at the URL (it clones shallow, read-only, and cleans up). Never run the skill's own scripts during vetting.
- Scan:
skillspector scan <dir> --recursive --no-llm --format json. The static scan needs no API key and no network beyond the clone. - Read
risk_assessment.severity(worst finding wins across a multi-skill repo):NONE/LOW-> adopt.MEDIUM-> read every issue and decide deliberately; record why if you adopt.HIGH/CRITICAL-> reject.
- Confirm it is actually a Claude skill: it must contain at least one
SKILL.md. A research framework or app with noSKILL.mdis not installable as a skill. - Record the verdict (repo, date, severity, decision) wherever the project tracks adopted assets, so the next person does not re-vet blind.
Notes
- Static scanning finds known-risky patterns, not every possible harm. For a MEDIUM result, human reading of the flagged lines is the gate, not the score alone.
- SkillSpector also has an LLM-assisted mode (drop
--no-llmwith a provider key) for deeper analysis; the static mode is the deterministic default for a quick gate.
