Source: raw/x-bookmark-2060610895458553977.md (X bookmark — @Dinosn / Nicolas Krassas, linking github.com/NVIDIA/SkillSpector; repo metadata verified via GitHub API 2026-05-31); install and --no-llm detail from raw/5_GitHub_Repos_-_Kill_AI_Slop_Go_Viral_Make_Money.md (Greg Isenberg, fetched 2026-09-29)
SkillSpector is an open-source security scanner for AI agent skills, published by NVIDIA. It statically inspects skill bundles for vulnerabilities, malicious patterns, and security risks before you install or run them — the skill-supply-chain analog of a dependency scanner. It addresses a real gap as the skills ecosystem grows: skills are executable instructions plus bundled files, and an untrusted skill can carry prompt-injection payloads, credential-exfiltration steps, or dangerous shell commands.
Key Takeaways
- What it does: scans AI agent skills for vulnerabilities, malicious patterns, and security risks (static analysis of skill bundles). ^[inferred — exact rule set / CLI flags not yet read from the README]
- Who ships it: NVIDIA (repo
NVIDIA/SkillSpector), Apache-2.0, Python. 483 stars, last pushed 2026-05-18 (at time of capture). - Why it matters: skills are run with the agent’s privileges. Vetting a third-party skill before enabling it is the skills-era equivalent of auditing an npm package — and the wiki’s skill-marketplace coverage (skills hubs, one-click installs) makes a vetting tool load-bearing.
- Fits the defense-in-depth theme already covered for Hermes (hermes-security-model) and the OSS-tool-gap-fillers for Claude Code (five-oss-tools-claude-code-blind-spots, which covers DeepSec) — SkillSpector targets the skill artifact specifically.
- Operator usage detail (2026-06 video demo). Invoked as
skillspector scan <repo-url>from a cloned repo + Python venv. The free run surfaces too many false positives to be useful; supplying an OpenAI API key sharpens it (cost ~$5 for a large repo, ~20–30¢ for a small one). In one demo it flagged a trending Chinese-language skill repo CRITICAL — 63 issues including cookie-pasting and RCE via unverified install scripts piped to the machine. This partially answers the CLI-usage Open Question below. (Source: raw/5_NEW_Vibe_Coding_Repos_You_Haven_t_Heard_of_Yet.md; see Vibe-Coding Tools Roundup.) - Install path and an offline mode (2026-09-29). Greg Isenberg, who says he has no NVIDIA affiliation, gives the quick install as
uv tool install git+<repo URL>, thenskillspector scan ./my-skillfor a local skill directory orskillspector scan <GitHub URL>for a repo. Adding--no-llmruns a faster static scan that does not send file contents to an LLM provider, which “matters if you’re scanning like sensitive files or private files.” That trades against the precision gain from supplying a model key noted above. A Docker option avoids installing Python locally. (Source: raw/5_GitHub_Repos_-_Kill_AI_Slop_Go_Viral_Make_Money.md.) - Run as a skill inside the agent (2026-09). Futurepedia installs SkillSpector as a callable skill. He invoked it in Cowork with the next skill’s repo link, and got a deep report rating humanizer safe to install. His calibration: it sometimes flags minor issues he installs past, and a clean report “isn’t an absolute guarantee.” The captions call it “Skill Inspector”. The repo,
NVIDIA/SkillSpector, had 22.7k stars on 2026-10-27, up from 483 at first capture. (Source: raw/9_Free_AI_Agent_Skills_You_NEED_to_Install_Now.md; see Nine Free Agent Skills.)
Try It
- Repo:
github.com/NVIDIA/SkillSpector(Apache-2.0, Python). - Run it against a skill you’re about to install from a public hub before enabling it — treat an unvetted skill like unvetted code.
- For a private or client skill, start with
skillspector scan ./my-skill --no-llmso nothing leaves your machine; add a model key only for public skills, where the sharper LLM pass is worth the upload.^[inferred — the private/public split is this wiki’s; the flag behaviour is from the source] - Pair with least-privilege install discipline (per-tool whitelisting, scoped credentials) as in hermes-mcp-catalog and hermes-security-model.
Related
- agent-skills-overview
- anthropic-skills-repo
- five-oss-tools-claude-code-blind-spots
- skills-ecosystem
- hermes-security-model
- Security-Guidance Plugin (Anthropic Official) — Anthropic’s first-party in-session scanner; SkillSpector vets skills you install, the security-guidance plugin vets code Claude writes.
- Claude Security Plugin — Anthropic’s first-party deep-scan layer; SkillSpector vets the skill supply chain, Claude Security scans the repository those skills operate on.
Open Questions
- Exact scan coverage (which vulnerability classes / injection patterns) and CI integration — needs a README + source pull (not yet deep-read; this article is from verified repo metadata + the source post). Basic CLI usage is now covered above (
scan,--no-llm, uv and Docker installs); how many false positives the--no-llmstatic pass produces is not reported. - Does it understand the Anthropic skill format specifically, or generic “agent skill” bundles? Verify before citing as Claude-skill-specific.