Summary
NVIDIA’s open-source SkillSpector is a security scanner built specifically for AI agent skills — the markdown-based instruction folders that give agents like Claude Code, Open Claw, and OpenAI Codex their capabilities. Research across tens of thousands of real-world skills found that roughly a quarter carry some form of vulnerability, and a smaller subset show signs of genuine malicious intent. Because skills execute inside an agent’s full trust context alongside user credentials, this represents a meaningful and largely unaddressed attack surface.
In this hands-on walkthrough, Fahd Mirza installs SkillSpector on Ubuntu with a single UV command and demonstrates two scans: a clean Git Committer skill (score: 0/100, severity: low, zero issues) and a modified malicious cooking-assistant skill (five issues flagged, severity: medium/caution). The scanner surfaces a hidden helper.py script — never mentioned in the skill’s markdown — that harvests environment variables and ships them to an external server. It also flags that the skill declares no tool scope, requesting unrestricted access by default.
SkillSpector’s pattern-based mode (–no-llm) covers 70-plus known attack patterns including prompt injection, credential theft, data exfiltration, and hidden instructions, and runs without an API key. For developers building or consuming agent skill ecosystems, this video provides both a clear threat model and a practical first line of defense before installing any third-party skill.
📺 Source: Fahd Mirza · Published September 06, 2026
🏷️ Format: Tutorial Demo
