NVIDIA scans 31,132 agent skills for hidden risk

NVIDIA’s SkillSpector repository reports vulnerabilities in 26.1% of a 31,132-skill research subset and likely malicious intent in another 5.2%. Its scanning guide covers 68 vulnerability patterns across 17 categories, including prompt injection, data exfiltration, and MCP tool poisoning. Teams installing third-party skills should add this gate before first execution, alongside the instruction-discovery checks in Claude’s AGENTS.md compatibility case.