SkillSpector can help you triage a skill before installing it. In our static-only test, all three synthetic samples came back SAFE, including one that quoted a fictional collection policy that went beyond its task and one with missing YAML frontmatter. Read the scan coverage and the skill together before granting access.
After static inspection, validate the tools called by a Skill separately. Inspect MCP responses with MCP Inspector and cross-check development usage through Context7's documentation lookup.
To build equivalent fixtures, create three folders with one SKILL.md each. Give the first two valid name and description fields. Put a summarization task in the first and an explicitly quoted fictional policy in the second. Leave off the YAML frontmatter in the third. Feed the files to the scanner as data.
Run a static scan and keep the JSON
Download the Python wheel from NVIDIA's v2.13.4 release, verify its listed SHA-256, and install it in an isolated Python environment. Run skillspector scan ./sample-skill --no-llm --format json > report.json. Read the verdict alongside issues and analysis_completeness.analyzer_statuses.
Compare three outputs
All three one-file fixtures exited with code 0 and reported zero issues. The plain summary skill and the fixture quoting a fictional broad collection policy both reported SAFE. The text was scanned as data and never run. The fixture without YAML frontmatter also reported SAFE, but its skill.name was unknown.
All three reports had risk_assessment.recommendation: SAFE, metadata.llm_requested: false, and analysis_completeness.is_complete: true. The semantic_developer_intent analyzer was listed as disabled. Check the individual analyzer statuses for disabled and inapplicable checks before interpreting overall completion.
Use the verdict to prioritize inspection
Check the skill's name and description, then map its requested access to the job you intend to give it. Read bundled scripts before allowing execution. If routine use asks for files, a network connection, or write permissions beyond the task, resolve that mismatch before installing.
Scope and upstream sources
We scanned local text in a network-disabled Linux container after installation. NVIDIA documents optional model analysis and OSV dependency checks. Enabling model analysis sends eligible file content to the configured provider. When online, the dependency check can send package names and versions to OSV even in static mode. The tool falls back to a bundled list when offline. This test did not call a model provider or verify live OSV results. SkillSpector is Apache-2.0 licensed. NVIDIA repository · Version 2.13.4 and release assets.