NVIDIA SkillSpector: security scanner for AI agent skills

I came across NVIDIA’s SkillSpector repo and it looks aimed at a very practical pain point: when you install third-party “skills” for agent runtimes, you’re effectively trusting code you didn’t write, often with elevated capabilities.

SkillSpector is positioned as a security scanner that inspects AI agent skills ahead of time to detect risky behavior patterns. The stated target surface includes Claude Code, Codex, and MCP skills, with checks for things like prompt injection, potential data exfiltration paths, and broader supply-chain risk signals.

When I would personally use this: suppose I’m wiring a small-team agent workflow (FastAPI service + LangGraph) and I want to add a community-provided MCP tool/skill. Before I even mount it into my runtime, I’d run SkillSpector against the skill package to see if it contains obvious injection primitives, suspicious exfiltration logic, or dependency patterns that look unsafe.

What to look at first: - The repo’s entry point / “how to run” instructions to understand what inputs it expects (skill manifest, package directory, or config). - The documented detections/flags so you can map alerts to concrete mitigations (e.g., sandboxing, allowlists, or stricter tool schemas). - Any examples in the README for Claude Code / Codex / MCP skill scanning workflows.

If you build or operate agent systems, this kind of pre-install scanning is a small but high-leverage step: it turns “trust the skill” into “verify the skill,” which is especially important when agent skills touch credentials, internal APIs, or external tools.