NVIDIA open-sources SkillSpector to scan Agent skills for security risks before install

SkillSpector accepts multiple input formats—Git repos, URLs, zip files, directories or single files—and ships 71 vulnerability patterns across 17 categories, covering prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, system prompt leakage, memory poisoning, tool misuse, MCP least privilege and tool poisoning. Analysis runs in two stages: fast static analysis plus optional LLM semantic evaluation, with real-time CVE lookups via SC4 against OSV.dev and automatic offline fallback. Output supports terminal, JSON, Markdown and SARIF, with a 0-100 risk score and severity labels; a baseline mechanism suppresses known false positives by fingerprint or glob rules, so repeat scans surface only new issues. It can also diff MCP Registry snapshots to list added, removed and changed server identities. For developers using Claude Code, Codex CLI, Gemini CLI and MCP skills, this provides a pre-install security gate—in the official subset of 31,132 skills analyzed, 26.1% contained vulnerabilities and 5.2% were suspected malicious. SARIF output plugs directly into CI/CD and IDEs, while baseline suppression lets teams freeze existing issues and alert only on new risks; --fail-on-incomplete can fail CI when scan coverage is incomplete. It is also part of the NVIDIA Verified Skills process, where passing skills are signed and published to the NVIDIA skills catalog. The repo includes Pi and OpenCode extensions for invoking scans directly inside Agent sessions; a Docker image runs without installing Python, and batch scanning scripts support Chinese, Japanese and Korean detection.