71 vulnerabilities divided into 17 categories, ranging from prompt injection to data exfiltration: the tool rates each vulnerability on a scale of 0 to 100 and advises against installation if the score exceeds 50.
NVIDIA has launched SkillSpector, an open source tool licensed under Apache 2.0 that analyzes "hands" for code agents — Claude Code, Codex CLI, Gemini CLI — as well as MCP servers, before they are installed. It accepts a Git repository, URL, zip archive, folder or simple file. The project cites work in which 26.1 per cent of the skills reviewed are vulnerable and 5.2 per cent are clearly malicious.
Static analysis, then semantic look
The scanner searches for 71 patterns grouped into 17 categories: quick injection, exfiltration of environmental variables, unpinned dependencies or obfuscated supply chain code, excessive agentivity, escalation of privileges, memory poisoning, too wide MCP permissions, etc. A first pass combines regular expressions, analysis of the Python syntax tree, followed by sensitive data flows and YARA signatures; dependencies are facing live OSV.dev. A second pass optionally entrusts an LLM to examine the context and intent, to rule out false positives.
The result takes the form of a risk score: low to 20, caution up to 50, and then "do not install" for high and critical levels. Reports come out in text, JSON, Markdown or SARIF, and the tool can itself run as an MCP server.
Read the full article on the original source
Go to the source →