SkillSpector
Security scanner published by NVIDIA for agent skills (Claude Code, Codex, Gemini CLI, MCP servers). You hand it a folder, a SKILL.md file, a Git repository or a zip archive before installing, and it looks for what a skill could do behind your back: prompt injection, data exfiltration, privilege escalation, vulnerable dependency, hidden content, a description that does not match the behaviour. 71 patterns across 17 categories, in two passes: a fast static analysis, then an optional review by a language model, the only pass that compares the stated intent with the code. The report comes out as terminal, JSON, Markdown or SARIF, with a 0 to 100 score. On a sample of 31,132 skills studied by the team, 26.1% had a vulnerability and 5.2% showed likely malicious intent. It is not an antivirus: it flags a risk, it neutralises nothing, and you are the one who decides not to install.