If you’re a beginner installing AI agent skills from the web, NVIDIA SkillSpector is a must-use safety net, 8/10 — here’s why. Imagine installing a “meeting notes” skill that quietly also reads your browser history or SSH keys in the background; that’s the kind of risk SkillSpector is built to catch before you ever hit install.
How We Tested
This review is based on official documentation, pricing pages, and published third-party reports — we did not run the tool hands-on. We covered setup and installation paths, the two analysis stages (static and optional LLM semantic), the 69 vulnerability patterns across 17 categories, the risk scoring system, the OSV.dev supply-chain lookup, output formats, and the underlying research study from the official NVIDIA scanning docs. We did not verify the LLM semantic pass’s ~87% precision claim hands-on; it is NVIDIA’s reported figure via Help Net Security.
| Overall Score | 8/10 | | Ease of Use | 8/10 | | Features | 9/10 | | Performance | 8/10 | | Documentation | 8/10 | | Support | 8/10 |
What Is SkillSpector?
SkillSpector is a free, open-source (Apache-2.0) security scanner built by NVIDIA that answers one question: “Is this skill safe to install?” before you add it to Claude Code, Codex CLI, Gemini CLI, or as an MCP server. The project sits at 14.6k stars on the SkillSpector GitHub repository with the current public OSS release at v2.9.3, and it serves as the security tier of NVIDIA’s Verified Skills pipeline.
Why Agent Skills Need a Security Scanner
Agent skills are essentially instructions — sometimes bundled with executable scripts — that run with your credentials and permissions. When you install a skill from a random GitHub repo or a marketplace, you’re granting that code access to your environment, your files, and potentially your API keys. The threat is not theoretical: the “Agent Skills in the Wild” study examined 42,447 skills and found that 26.1% contain at least one vulnerability, 5.2% show likely malicious intent, and skills with executable scripts are 2.12x more likely to be vulnerable. In plain terms, roughly 1 in 4 skills has a problem.
Here’s a concrete scenario, framed purely as illustrative: imagine a “summarize PDFs” skill. Its visible instruction says it reads PDFs and outputs summaries. But the bundled Python script, hidden in a subdirectory, also iterates through your ~/.ssh/ directory and sends the contents to a remote server. Nothing about that behavior is visible in the skill’s description — which is exactly why a scanner like SkillSpector exists.
How SkillSpector Works
SkillSpector uses a two-stage analysis pipeline. The first stage is a fast static scan that runs in seconds and requires no API key — it examines the skill’s code and structure for known dangerous patterns. The second, optional stage is an LLM semantic analysis that requires an OpenAI-compatible provider (configured via the SKILLSPECTOR_PROVIDER environment variable, defaulting to NVIDIA’s build.nvidia.com); this stage catches mismatches between what a skill claims to do and what its code actually does.
The static pass checks 69 vulnerability patterns across 17 categories. A few worth knowing in plain language:
- Prompt injection — tricking the AI into ignoring its instructions
- Data exfiltration — stealing files or secrets
- Privilege escalation — grabbing more access than intended
- Supply chain — compromised dependencies
- MCP tool poisoning — manipulating the tools the AI uses
After scanning, SkillSpector computes a risk score from 0 to 100. Each finding adds points; a score above 50 means “do not install.” If the skill contains executable content, the score is multiplied by 1.3x — reflecting the study’s finding that scripts significantly increase vulnerability risk. The tool also performs a live CVE lookup via OSV.dev, batching the package list into a single query with a 1-hour cache and an offline static fallback. Results can be output to terminal, JSON, Markdown, or SARIF (a standard format CI systems use for security findings).
How Beginners Use It
Getting started is straightforward. Install via uv tool install git+https://github.com/NVIDIA/skillspector.git, or use Docker if you don’t want Python 3.12+ on your machine. Then run:
skillspector scan ./my-skill/The output shows the risk score and a list of findings. If you want a static-only check without configuring an LLM provider, add --no-llm. Use --format json or --format sarif to pipe results into other tools. The full setup and analysis details are in NVIDIA’s official scanning docs.
Alternatives to Consider
- Deno Claw Patrol — a runtime firewall that guards agent actions at execution time instead of scanning before install. It’s complementary, not identical — you could use both.
- agentskills.io skills-ref — lightweight spec validation, not a security scanner; it checks format, not malicious behavior.
- General LLM red-teaming tools (Promptfoo, Garak) — broader AI-security scope, not focused on skill packages specifically.
See how these fit in our AI Agent Skill comparison, or browse the full comparisons hub.
Beginner Takeaway
If you’re new to agent skills, start with SkillSpector because a 30-second scan before every install is the cheapest insurance you’ll ever buy. It’s free, it’s fast, and it directly addresses the 1-in-4 odds that a skill has a vulnerability. For related reading, see our Ponytail skill review and our Impeccable skill review to understand what good skills look like.
Roadmap Hook
This review is part of our AI Agent Skill category expansion — we’re at 2/5 tools and adding more weekly. Track progress at /roadmap/.
Who Should Use This / Who Shouldn’t
- Who should: anyone who installs skills from third-party repos or marketplaces — Claude Code, Codex CLI, Gemini CLI, MCP servers — and wants a fast pre-install safety check.
- Who shouldn’t: users who only install first-party or whitelisted skills; people who need runtime monitoring (Claw Patrol is the better fit); environments where Python 3.12+ or Docker isn’t feasible.
See what’s next in our review pipeline at /cron-pipeline/.
📖 Related Reads
- Hermes Tutorials — Hermes Agent setup, configuration, and advanced workflows
- NiteAgent — AI agent development, frameworks, and production patterns
- ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
Cross-links automatically generated from None.
Back to all posts
