Repolode
Submit
NVIDIA

NVIDIA/SkillSpector

Security scanner that checks AI agent skills (Claude Code, Codex, MCP) for prompt injection, data exfiltration and supply-chain risks before you install them.

View on GitHubWebsite
Stars
20,912
Forks
1,836
Open issues
170
Last commit
1 hour ago

What is it?

AI agent skills run with implicit trust, and SkillSpector is a scanner that inspects them before installation. It accepts Git repositories, URLs, zip archives, directories or single SKILL.md files, runs fast static analysis across dozens of vulnerability patterns, and can optionally add LLM-based semantic review through a range of providers including OpenAI, Anthropic, AWS Bedrock, Ollama and any OpenAI-compatible endpoint.

Results are reported as a 0-100 risk score with a severity level and a recommendation such as SAFE, CAUTION or DO NOT INSTALL. Output can be terminal text, JSON, Markdown or SARIF for CI pipelines. A baseline file lets teams accept known findings so that re-scans only surface new issues.

SkillSpector can also run as an MCP server so that agents such as Claude Code can call the scanner as a tool and gate installs on the result.

Key features

  • 01Scans local directories, single files, zip archives, URLs and Git repositories
  • 02Detects prompt injection, data exfiltration, credential access and supply-chain issues
  • 03Combines static analysis with optional LLM semantic evaluation
  • 04Outputs terminal, JSON, Markdown and SARIF reports
  • 05Supports baseline files to suppress accepted findings on re-scan
  • 06Runs as an MCP server for agent-driven install gating

Use cases

  • Vetting a third-party Claude Code or Codex skill before installing it
  • Gating agent skill installs in a CI pipeline using SARIF output
  • Reviewing MCP server definitions for excessive permissions or tool poisoning
  • Scanning a collection of skills in batch to triage risk

Quick start

  1. 1

    Install SkillSpector as a CLI tool with uv

    uv tool install git+https://github.com/NVIDIA/skillspector.git
  2. 2

    Scan a local skill directory without LLM analysis

    skillspector scan ./my-skill/ --no-llm

Things to know

  • Full semantic analysis requires configuring an LLM provider with its own API key or endpoint.
  • Installing from Git requires cloning the repository and fetching third-party open source dependencies.

Topics

#agent-security#ai-security#prompt-injection#supply-chain-security#mcp#claude-code#security-scanner#agent-skills#agentic-ai#security-tools

FAQ

What is SkillSpector?
Security scanner that checks AI agent skills (Claude Code, Codex, MCP) for prompt injection, data exfiltration and supply-chain risks before you install them. AI agent skills run with implicit trust, and SkillSpector is a scanner that inspects them before installation. It accepts Git repositories, URLs, zip archives, directories or single SKILL.md files, runs fast static analysis across dozens of vulnerability patterns, and can optionally add LLM-based semantic review through a range of providers including OpenAI, Anthropic, AWS Bedrock, Ollama and any OpenAI-compatible endpoint.
Is SkillSpector free to use? What license does it have?
SkillSpector is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is SkillSpector still maintained?
Actively maintained. The last commit to SkillSpector was on Oct 22, 2026.
What can I use SkillSpector for?
Vetting a third-party Claude Code or Codex skill before installing it Gating agent skill installs in a CI pipeline using SARIF output Reviewing MCP server definitions for excessive permissions or tool poisoning Scanning a collection of skills in batch to triage risk
How do I install SkillSpector?
Install SkillSpector as a CLI tool with uv: uv tool install git+https://github.com/NVIDIA/skillspector.git Scan a local skill directory without LLM analysis: skillspector scan ./my-skill/ --no-llm
Where can I get SkillSpector?
The source code and installation instructions are on GitHub: https://github.com/NVIDIA/SkillSpector

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects