Verified project record

NVIDIA/SkillSpector

skillspector is a security scanning tool that helps users determine whether AI agent skills are safe to install by detecting vulnerabilities, malicious patterns, and security risks before installation. It scans multiple input formats and identifies dozens of vulnerability patterns across distinct categories using static analysis with optional LLM-based semantic evaluation.

Project overview

As AI agent skills are routinely installed with implicit trust, skillspector addresses an emerging security gap by providing pre-installation vetting through multi-format scanning and a two-stage analysis pipeline.

Project type: MCP · AI Agent · Evaluation & Observability
Use cases: Automation
Deployment: Refer to project documentation
License: Apache-2.0

Best for

Developers and AI engineers who need to evaluate the security of AI agent skills before installation or deployment.
Operations teams seeking programmatic or CLI-based scanning with structured reporting for automated pipelines.

Key capabilities

Scans Git repositories, URLs, zip files, directories, or individual files.
Detects 68 vulnerability patterns across 17 categories including prompt injection, data exfiltration, and privilege escalation.
Performs fast static analysis alongside optional LLM semantic evaluation.
Queries OSV.dev for real-time CVE data with an automatic offline fallback.
Generates reports in Terminal, JSON, Markdown, and SARIF formats.
Provides a 0-100 score accompanied by severity labels and clear recommendations.
Suppresses known findings via glob-rule or fingerprint baseline so that re-scans surface only new issues.
Can be installed with the MCP extra to run the `skillspector mcp` command.

Limitations and risks

The tool relies entirely on static analysis and does not perform dynamic execution or runtime behavior evaluation.
Security patterns embedded within non-English content may be missed during analysis.
Cannot analyze or extract text concealed within images.
Cannot analyze compiled, encrypted, or otherwise obfuscated binary content.
Without network access to api.osv.dev, live vulnerability lookups fall back to a small static list, which may limit real-time CVE coverage.

Getting started

Install easily using `uv tool install git+https://github.com/NVIDIA/skillspector.git` and immediately scan a skill directory by running `skillspector scan ./my-skill/`.
Users can run the scanner via Docker without needing to install Python directly on their system.

Evidence and sources

GitHub project description: Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, and security risks.
README: AI agent skills (used by Claude Code, Codex CLI, Gemini CLI, etc.) execute with implicit trust and minimal vetting. Research shows that **26.1% of skills contain vulnerabilities**…
README: **[Pi extension](docs/PI_EXTENSION.md)** — Install SkillSpector as a Pi tool for scanning skills from inside agent sessions.
README: - **Multi-format input**: Scan Git repos, URLs, zip files, directories, or single files
README: - **68 vulnerability patterns** across 17 categories: prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, output handling, system prompt lea…

Reviewed at: 2026-07-17T23:57:40Z