AI agent skill files (SKILL.md, AGENTS.md, MCP configs, custom instructions) sit inside the engineering trust boundary and introduce supply-chain risk. NVIDIA SkillSpector can scan these files for prompt-injection patterns, exfiltration behavior, unsafe tool-usage chains, and policy-bypass tactics. Two GitHub Actions workflows are provided: a baseline repository scan that produces SARIF output uploaded to GitHub Code Scanning, and a focused PR scan that identifies changed skill directories, scans only those, and posts results as a PR comment. The post covers exit code handling, LLM-backed semantic analysis via Azure AI Foundry, artifact retention, and practical guidance on when to use targeted vs. full-repo scanning.
387 Impressions