SkillSpector

What it is

An NVIDIA project that answers one question before you add a skill to Claude Code, Codex CLI, Gemini CLI, or an MCP setup: is this skill safe to install? It scans Git repos, URLs, zip files, directories, or a single SKILL.md, and gives a 0 to 100 risk score with severity labels and recommendations.

What it helps with

- Checking third-party agent skills against 71 vulnerability patterns in 17 categories, including tool misuse, memory poisoning, and MCP tool poisoning
- Fast static analysis, with an optional LLM pass for semantic review
- Looking up known vulnerabilities in dependencies through OSV.dev, with an offline fallback
- Reporting in terminal, JSON, Markdown, or SARIF output
- Suppressing accepted findings with a baseline so re-scans show only new issues

Who it's for

Developers, platform teams, and security reviewers who install or publish agent skills.

Worth knowing

It is a Python 3.12+ command-line tool and can also run in Docker. NVIDIA uses it in its Verified Skills pipeline, which scans and signs skills before they reach the NVIDIA skills catalog. The LLM stage is optional and sends skill content to the model provider you configure.

Can I run it without sending code to an LLM?
Yes, static-only scans run with the --no-llm flag.

How does it differ from garak or PyRIT?
In our view they complement each other: garak and PyRIT probe models for weaknesses, while SkillSpector reviews the skills an agent is about to load.