SkillSpector: A scanner that detects malicious patterns in AI agent skills before installation

AI agent skills such as Claude Code, Codex, and MCP are often executed without sufficient verification at the time of installation. NVIDIA SkillSpector is a security scanner that detects vulnerabilities and malicious patterns within these skills in advance. It provides safety verification based on the fact that, out of 31,132 skills analyzed in a research dataset, 26.1% exhibited vulnerabilities and 5.2% showed malicious intent.

It supports various input formats, including Git repositories, URLs, zip files, directories, and single files. It identifies 71 types of vulnerability patterns across 17 categories, including prompt injection, data leakage, privilege escalation, and supply chain attacks. It undergoes a two-stage analysis process consisting of fast static analysis and optional LLM-based semantic evaluation, and queries real-time CVE data through OSV.dev.

Beyond simple detection, it assigns a risk score from 0 to 100 and severity labels to indicate remediation directions. It outputs reports in various formats such as terminal, JSON, Markdown, and SARIF, making it suitable for integration with CI/CD pipelines or IDE tools. It also provides a feature to register known detection results as a baseline, displaying only new issues during rescans.

As part of the NVIDIA Verified Skills pipeline, skills that pass are published in the NVIDIA skill catalog. Developers who deploy agent skills or adopt third-party skills can use this tool to check security risks before installation. It is released under the Apache-2.0 license and can be run without a Python environment via a Docker image.