SkillSpector is NVIDIA’s Apache-2.0 security scanner for AI agent skills: point it at a git repo, URL, zip, directory, or a single SKILL.md, and it returns a 0-100 risk score with a SAFE, CAUTION, or DO NOT INSTALL recommendation before you install the thing.
Its bet is that skills execute with implicit trust and minimal vetting, and it opens with its own research numbers: in a 31,132-skill analyzed subset, 26.1% of skills contained vulnerabilities and 5.2% showed likely malicious intent.
What it is #
A Python 3.12+ CLI (installed with uv or pip from git, or run as a Docker image) from NVIDIA, built as two analysis stages.
Stage one is static: eleven analyzers covering 71 vulnerability patterns across 17 categories, prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, MCP tool poisoning and least privilege, plus AST behavioral checks, taint tracking, YARA signatures, and live CVE lookups against OSV.dev with an offline fallback.
Stage two is an optional LLM semantic pass (OpenAI, Anthropic, Bedrock, Gemini, Ollama, local runtimes, or NVIDIA’s own build.nvidia.com endpoint) that can confirm findings and add explanations but never removes one, and disputed findings stay tagged llm-unconfirmed in the JSON and SARIF output.
Reports land as terminal text, JSON, Markdown, or SARIF for CI, the exit codes are a documented install-gate contract, and an MCP server mode exposes a single scan_skill tool so an MCP host can gate installs at runtime.
It is the security tier inside NVIDIA’s SkillEvaluator and part of the Verified Skills pipeline that scans, evaluates, and signs skills before publishing them to NVIDIA’s catalog, with in-session extensions for Pi and OpenCode.
Status #
Active and large: 20,139 stars and 1,770 forks as of 2026-10-13 (1,857 forks as of 2026-10-24, 2,359 as of 2026-11-17), 163 open issues as of 2026-11-19, created 2026-03-21, main’s last commit 2026-11-19 (parser-boundary and command-matching fixes: command boundaries and printf braces, reconstructed Python operands now requiring viable options, runtime command bindings, PRs #1151 through #1154), v2.13.6 the newest release.
The release train is active: v2.13.6 is the latest release, published 2026-11-16 (complete JSON values such as {"name":"nice"} no longer produce false static_parse_limit or command_word_boundary diagnostics, with command arguments and unresolved fragments still checked, plus documentation updates on --min-coverage thresholds, JSON archive-member provenance, and least-privilege remediation guidance), published hours after v2.13.5 (2026-11-16, improved bounded parsing and inspection of large skills and structured data with explicit exhausted-limit diagnostics, preserving the provider defaults, output-file protections, and passive Office metadata handling of 2.13.4), which itself follows v2.13.4 (published 2026-11-04, provider compatibility and structured responses for GPT-6.1 Sol and Claude Opus 5.5, bounded parsing of shell, PowerShell, JavaScript template, Python string, and Markdown sources, no-execution inspection of generated shell helpers and archived sources, bounded TAR/GZIP inspection with explicit omitted-member accounting, and provider-failure detail preservation with bounded retries) and v2.13.3 (2026-11-02, selected P0 security and scan-integrity fixes including four dependency security updates, a configurable tool-description batch limit defaulting to 256, and a workflow deadline raised to 1,080 seconds), preserving the dependency lock and provider defaults; it follows v2.13.1 (2026-10-23, false-positive fixes for incomplete-analysis reports, quoted text, and HTML-comment-spanning hidden-instruction findings, three backported security fixes covering report-file writing, failed batch gap-fill reporting, and bounded tool-misuse context), v2.13.0 (2026-10-16, a security-hardening backport of 55 selected PRs onto v2.12.0 (2026-09-23, whose notes close with “Prepared by Codex on behalf of Mohit Gupta”)), which pins the loopback-only MCP HTTP restriction and the disabled Codex CLI provider into the release contract itself.
The Hacker News footprint is one 49-point thread with 4 comments from 2026-06-12, thin against the star count and the same stars-ahead-of-discussion pattern this section has flagged on agentmemory and Cabinet.
The project’s documentation already disagrees with itself: the README counts 71 vulnerability patterns, the hosted docs guide counts 68.
Strengths #
- Fail-closed engineering is rare in this ecosystem: ingest caps (100 MiB per source, 10,000 zip members), coverage accounting where any excluded file blocks a SAFE verdict, and an exit-code contract built for CI gates.
- Deterministic findings are immovable: the LLM stage can confirm or dispute, but every static finding stays in the report.
- Baseline suppression lets a team accept known findings once and resurface only new ones, with evidence-bound fingerprints that reactivate when the source changes.
- The README states its own limits, including moderate static precision, no guaranteed LLM precision, and a disabled codex_cli provider whose sandbox permits host-file reads.
Cautions #
- The pattern set is one vendor’s opinion, and the 68-versus-71 count drift between the README and the docs guide shows the taxonomy is still moving.
- The OpenCode in-session provider pins one exact OpenCode version (1.18.33 in the README, 1.18.31 in the v2.12.0 notes), so in-session semantic scans break on any other OpenCode release.
- LLM-stage scans bill your own provider, and the batch scanner’s default model targeted DeepSeek, which the README expects to sunset.
- The MCP HTTP transport ships without authentication by design (loopback-only binding, credential-free https targets only), a documented but easy-to-misuse trust model.
- A clean scan inspects the artifact pre-install; it says nothing about whether the skill improves your output, which is SkillEvaluator’s separate question.
Pricing #
Free and Apache-2.0, with no hosted tier of its own. Static scans cost nothing; LLM-stage scans run on your own provider credentials and bill your tokens.
Compared to #
- SkillMD: the registry that publishes lint and scanner verdicts per indexed skill; SkillSpector is the scanner you run yourself on a skill SkillMD has not reviewed.
- skills.sh: the registry whose audit columns aggregate third-party scanner output after the fact; SkillSpector is the gate run before install rather than a column read after.
- OpenCode skills and plugins: the harness-side allow/deny/ask gate at load time; SkillSpector inspects content before it ever reaches a harness, so the two compose rather than compete.
Bottom line #
Recommended for teams that install third-party skills into CI or shared images, where the exit codes and SARIF output make a pre-install scan a cheap gate, and for anyone about to run an unvetted skill against a repo with production secrets. Not for runtime policy enforcement (that is the harness or a control plane), for output-quality questions (that is SkillEvaluator’s separate tier), or as a clean bill of health, since its static stage advertises moderate precision and the LLM stage guarantees nothing removed.
Changes #
- 2026-10-12 - Created from the daily-refresh scan resolution, with five fetched sources, the one-vendor-pattern-set caution, and the docs-versus-README pattern-count drift recorded as the critical angle.
- 2026-10-17 - The release train moved to v2.13.0 (published 2026-10-16), a security-hardening backport of 55 selected PRs onto v2.12.0 that pins the loopback-only MCP HTTP restriction and the disabled Codex CLI provider into the release contract; the Cautions on both now rest on release notes as well as the README.
- 2026-10-24 - The release train moved to v2.13.1 (published 2026-10-23), a patch on the v2.13.0 maintenance branch carrying false-positive fixes and three backported security fixes (atomic report-file writing, failed batch gap-fill passes now reported as incomplete with exit code 2, bounded tool-misuse context) while preserving the dependency lock and provider defaults.
- 2026-10-27 - Volatile refresh: the release train held at v2.13.1 with main’s last commit still 2026-10-22 (the pushed date moved to 2026-10-27 on non-main activity) and open issues up from 164 to 181.
- 2026-10-31 - The release train moved to v2.13.2 (published 2026-10-28), a patch on the v2.13.1 maintenance branch backporting three false-incomplete-analysis fixes (wrapper commands like timeout, sudo, nice, and xargs at proven clause boundaries no longer trip parser-limit reports) while preserving the dependency lock and provider defaults; main pushed again to 2026-10-30 with scanner false-positive and LLM-batching fixes, and open issues moved down to 147.
- 2026-11-07 - The release train moved twice more: v2.13.3 (published 2026-11-02, selected P0 security and scan-integrity fixes with four dependency security updates, a configurable tool-description batch limit defaulting to 256, and the workflow deadline raised to 1,080 seconds) and v2.13.4 (published 2026-11-04, provider compatibility for GPT-6.1 Sol and Claude Opus 5.5, bounded parsing of shell, PowerShell, JavaScript template, Python string, and Markdown sources, no-execution inspection of generated helpers and archived sources, and bounded TAR/GZIP inspection); main pushed again to 2026-11-06 (contributor-onboarding verification docs plus parser-boundary, LLM-batch-admission, and refusal-accounting fixes), and open issues moved to 171.
- 2026-11-12 - Volatile refresh: main pushed to 2026-11-11 with five merged PRs (generated-shell-source inspection, bounded static pattern matching, workspace alias checks preserved, restricted inherited report access, credential access detection), the pushed date’s 11-12 move landing on side-branch activity, and open issues moved down to 158; v2.13.4 still the newest release.
- 2026-11-13 - Volatile refresh: main pushed again to 2026-11-13 (one fix keeping npm link slots from borrowing ancestor pins) and open issues moved down to 150; v2.13.4 still the newest release.
- 2026-11-14 - Volatile refresh: main pushed to 2026-11-14 with five merged PRs since the 11-13 reading (bare JSON wrapper-name completion, MiniMax regional support with M3.1 Flash defaults, ignore-list code fences in Markdown treated as ignore-file entries, perseus-ctx and mimir-mcp removed from the popular-PyPI package list, the nv_build default model switched to z-ai/glm-5.3-flash), and open issues moved down to 145; v2.13.4 still the newest release.
- 2026-11-17 - Volatile refresh: the release train moved twice on 2026-11-16, v2.13.5 (improved bounded parsing and inspection of large skills and structured data with explicit exhausted-limit diagnostics, preserving the 2.13.4 provider defaults and output-file protections) and v2.13.6 hours later (complete JSON metadata with bare wrapper names like
{"name":"nice"}no longer trips false static_parse_limit or command_word_boundary diagnostics, plus documentation updates on –min-coverage thresholds, JSON archive-member provenance, and least-privilege remediation guidance); main pushed again to 2026-11-16 (README documentation links, CLI import deferral for help and registry commands, an MCP-registry UTF-8 BOM fix), and open issues moved up to 157. - 2026-11-19 - Volatile refresh: main pushed to 2026-11-19 with the parser-boundary batch (command boundaries and printf braces, reconstructed Python operands requiring viable options, runtime command bindings, PRs #1151 through #1154) and open issues moved up to 163; v2.13.6 still the newest release.
See also #
- SkillMD - the verification-first registry, the at-scale version of the per-scan verdict SkillSpector computes
- skills.sh - the registry whose audit columns aggregate scanner output, the market SkillSpector’s verdicts feed
- SkillOpt - the other tool that operates on skill artifacts as inputs, training rather than grading
- Anthropic Agent Skills - the vendor framing of skills as untrusted inputs that motivates pre-install scanning
- Agent Skills open standard - the format whose six fields carry none of the security posture this scanner checks
References #
https://api.github.com/repos/NVIDIA/SkillSpector - 20,139 stars and 1,770 forks as of 2026-10-13 (1,857 forks as of 2026-10-24), 171 open issues as of 2026-11-07, 158 as of 2026-11-12, 150 as of 2026-11-13, 145 as of 2026-11-14, 157 as of 2026-11-17, and 163 as of 2026-11-19, Apache-2.0, created 2026-03-21, main’s last commit 2026-11-19, docs homepage (fetched 2026-11-07, re-fetched 2026-11-12, 2026-11-13, 2026-11-14, 2026-11-17, and 2026-11-19)
https://raw.githubusercontent.com/NVIDIA/SkillSpector/main/README.md - the 71-pattern taxonomy, two-stage pipeline, MCP mode, install paths, risk scoring, and the 31,132-skill research numbers (fetched 2026-10-22, the 71 count and the OpenCode 1.18.33 pin unchanged)
https://docs.nvidia.com/skills/scanning-agent-skills - the hosted guide with its 68-pattern count, the triage policy, and the SkillEvaluator security-tier framing (fetched 2026-10-12)
https://api.github.com/repos/NVIDIA/SkillSpector/releases/latest - v2.13.6 published 2026-11-16 (JSON-metadata bare-wrapper-name false-diagnostic fix plus documentation updates, wheel and source distribution, no PyPI publish), following v2.13.5 published 2026-11-16 (bounded-parsing and inspection improvements preserving the 2.13.4 provider defaults), v2.13.4 published 2026-11-04 (GPT-6.1 Sol and Claude Opus 5.5 provider compatibility, bounded parsing, no-execution inspection, bounded TAR/GZIP), and v2.13.3 published 2026-11-02 (P0 security and scan-integrity fixes, four dependency updates, the 18-minute workflow deadline), and v2.13.0 published 2026-10-16, a security-hardening backport of 55 selected PRs onto v2.12.0 (sha256-digested assets, Codex-prepared notes, published 2026-09-23) (fetched 2026-11-07, re-fetched 2026-11-17)
https://hn.algolia.com/api/v1/search?query=SkillSpector&tags=story - the single 49-point, 4-comment thread (2026-06-12) behind the thin-footprint statement (fetched 2026-10-12)