What happened
NVIDIA released SkillSpector (Apache-2.0, on GitHub), a scanner for Claude Code, Codex CLI and Gemini CLI agent skills that catches prompt injection, data exfiltration, privilege escalation, supply-chain risk, memory poisoning, MCP least-privilege and tool-poisoning issues across 71 vulnerability patterns in 17 categories, with two-stage static-plus-LLM analysis, live OSV.dev CVE lookups, SARIF/Markdown output, and a 0-100 risk score. Its research subset found vulnerabilities in 26.1% of 31,132 skills and likely malicious intent in 5.2%.
Why it matters
Agent skills install with implicit trust and minimal vetting; a vendor-neutral, open scanner that gates skill installation addresses one of the highest-value, least-covered attack surfaces of the agentic era, and is part of NVIDIA's Verified Skills signing pipeline.
Applicability
Teams adopting Claude Code, Codex or Gemini skills (or publishing skills) should add SkillSpector to install/pre-commit gates and CI, and use it alongside MCP scanning for runtime supply-chain safety.