NVIDIA SkillSpector scans MCP skills for prompt injection before runtime

NVIDIA released SkillSpector, a static analyzer that inspects MCP skill definitions for prompt injection, data exfiltration patterns and suspicious dependencies before tools reach an agent runtime. It plugs into CI and blocks merges on critical findings.

• SkillSpector scans MCP tool descriptions and parameter schemas for hidden instructions
• The --fail-on-critical flag blocks pull requests on critical findings
• It cannot catch attacks assembled at runtime from environment variables
• Offline scanning needs pre-cached npm registries