NVIDIA SkillSpector
Security scanner for AI agent skills. Detects vulnerabilities, malicious patterns, and supply-chain risks before a skill is installed. Apache 2.0, ~11.6K stars. Source: NVIDIA/SkillSpector README, GitHub
SkillSpector answers one question: "Is this skill safe to install?" It exists because agent skills used by Claude Code, Codex CLI, and Gemini CLI execute with implicit trust and minimal vetting. NVIDIA cites research that 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent. Source: NVIDIA/SkillSpector README, GitHub
How It Works
- 64 vulnerability patterns across 16 categories: prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, output handling, system-prompt leakage, memory poisoning, tool misuse, rogue agent, trigger abuse, dangerous code (AST), taint tracking, YARA signatures, MCP least privilege, and MCP tool poisoning.
- Two-stage analysis: fast static analysis plus optional LLM semantic evaluation (OpenAI, Anthropic, or NVIDIA
build.nvidia.comproviders, or any local OpenAI-compatible endpoint). - Live CVE lookups: queries OSV.dev for supply-chain checks with offline fallback.
- Multi-format input: scans Git repos, URLs, zip files, directories, or single
SKILL.mdfiles. - Risk scoring: 0-100 score with severity labels, output to terminal, JSON, Markdown, or SARIF for CI/CD and IDE tooling.
Source: NVIDIA/SkillSpector README, GitHub
skillspector scan https://github.com/user/my-skill --format sarif --output report.sarif
Where It Fits Kevin's Stack
SkillSpector slots into the three-gate quality chain alongside Security and Review Skills (6-step manual vetting) and Brin - Agent Context Security (context-level allowlist scoring). The SARIF output and GitHub Action wiring make it the CI-automatable member of that chain: gate every community-submitted skill before it goes live, rather than vetting by hand. Daniel San (@dani_avila7) is adding it as a GitHub Action to his aitmpl.com template catalog so every submitted skill is scanned before publication. Source: X/@dani_avila7, 2026-05-31
A genuine install candidate (NVIDIA-published, real README, ~11.6K stars). Run the three-gate chain (Brin, skill-auditor, reputation) before adopting, per AGENTS.md.
Version Snapshot
As of 2026-06-30, NVIDIA/SkillSpector has no public release tags; use source commit snapshots instead. Current default-branch HEAD is 5df93c5b4c3ed6e80dd8ba76c8645cde484e4e30, license Apache-2.0, pushed 2026-06-30, with about 11.6K GitHub stars. This should be treated as a source-based security tool until NVIDIA publishes stable releases or a package channel. Source: GitHub API + git ls-remote, 2026-06-30
Kevin stack decision: add SkillSpector to the Security And Supply Chain capsule and keep it as the CI-automatable skill-supply-chain scanner. It complements Security and Review Skills rather than replacing it: SkillSpector catches static/semantic risk classes at scale; skill-auditor still owns human judgment about whether the skill should enter the active stack.
Timeline
- 2026-06-30 | Deep-reviewed the SkillSpector bookmark and current repo metadata. Updated reputation from ~928 to ~11.6K stars, recorded no release tags / HEAD
5df93c5b4c3ed6e80dd8ba76c8645cde484e4e30, and promoted it into the security capsule as the CI-automatable skill scanner. Source: X/@dani_avila7, 2026-05-31; Source: GitHub, 2026-06-30 - 2026-05-31 | @dani_avila7 surfaced SkillSpector as exactly the tool needed to secure agent skills, planning a GitHub Action integration with aitmpl.com (498 likes, 853 bookmarks). Source: X/@dani_avila7, 2026-05-31