SkillSpector

Agent skills run with a lot of trust and almost no review. SkillSpector is NVIDIA’s open scanner for that surface: point it at a git repo, zip, URL, or folder and it scores whether the skill is safe to install. Their research slice found over a quarter of skills with vulnerabilities and a few percent with likely malicious intent. It sits in the NVIDIA Verified Skills pipeline that scans, evaluates, and signs packs before catalog publish.

Founders are installing skill dumps and MCP servers into Claude Code, Codex, and Cursor every week. One bad skill can leak keys or widen agency. Deepsec-style repo audits catch bugs in your app; this catches the plugins you bolt onto the agent before they touch production.

Install the CLI (or Docker), then run skillspector scan on a path or URL. Fast static analysis covers 71 patterns across 17 classes, with optional LLM semantic review and live OSV lookups. You get a 0–100 risk score plus terminal, JSON, Markdown, or SARIF output. With the MCP extra it exposes scan_skill so an agent can gate installs at runtime. Baselines suppress known noise so re-scans only show new findings.

Most agent security tools hunt bugs in your product code. SkillSpector is built for the skill and MCP supply chain itself—prompt injection, tool poisoning, exfil patterns, and least-privilege checks—plus an install-time MCP gate, not a one-off report.

• MCP server / client support