Securing AI Agent Skills with NVIDIA Skill Spector

Detecting Malicious Agent Skills

AI agent skills are often distributed as text files containing instructions that agents execute with high trust. Research indicates that over 25% of these skills contain vulnerabilities, ranging from credential theft to malware execution. NVIDIA's Skill Spector tool addresses this by scanning skills for 14 common attack vectors, categorized into six primary threats: hidden instructions (obfuscated code or invisible characters), tool impersonation (using lookalike characters to spoof trusted tool names), deceptive descriptions (where code behavior contradicts documentation), credential harvesting, malware injection (such as reverse shells), and poisoned dependencies (typosquatted packages).

Automated Security Workflows

Skill Spector operates in two modes. The first is a pattern-matching scan that identifies known malicious signatures and suspicious file structures. The second mode utilizes an LLM to analyze the intent of the code, which is necessary to catch deceptive skills that pass static analysis. While the AI-based scan typically requires an OpenAI API key, users can bypass this cost by leveraging Claude Code's headless mode to execute the analysis using Anthropic's infrastructure. By wrapping Skill Spector into a custom 'discovery' skill, developers can automate a secure workflow: searching for new skills via repositories like skills.sh, scanning them for threats, and only proceeding with installation if the security score meets safety thresholds.