If you're using AI coding agents and assistants, you've probably come across what are called 'AI skills.' These aren't human skills; they're small, installable software packages that help AI agents like Claude Code or Gemini CLI perform specific tasks. At first glance, these skills seem totally harmless—just a text file or a few simple instructions. But the truth is, they behave like mini-software, and any software needs careful security review. The problem is that the risks in an AI skill aren't always obvious. They might not look like traditional malware or have scary filenames. Risks can hide in places that are easy to overlook, such as instructions in Markdown files, prompt injection text, helper scripts, attempts to access environment variables, self-modifying behaviors, overbroad permissions, or suspicious external dependencies. Traditional security review asks: 'What code will run?' But AI skill review also needs to ask: 'What behavior is this skill trying to influence?' These are related but not the same question. OWASP, well-known for web security, already lists risks like prompt injection, sensitive information disclosure, and supply-chain vulnerabilities in its Top 10 for LLM applications. NVIDIA's SkillSpector project highlights this further; research showed that 26.1% of skills contained vulnerabilities, and 5.2% showed likely malicious intent. These numbers are quite worrying! If AI skills are easy to install and share but can be risky, then simply 'installing it and seeing what happens' isn't a safe review process. This is where NVIDIA SkillSpector comes in. It's a security scanner for AI agent skills that helps answer a simple but crucial question: 'Is this skill safe to install?' The tool can scan folders, files, zip archives, and even repositories, looking for risky patterns before they can cause any harm. This way, you can harness the power of AI without worrying about hidden dangers.