Formal Analysis and Supply Chain Security for Agentic AI Skills

arXiv 2603.00195·high signal

Large-scale empirical scan of 42,447 agent skills across multiple registries finds 26.1% exhibit at least one security vulnerability across 14 distinct vulnerability patterns, including MCP-ITP (automated implicit tool poisoning) and MCPTox benchmark attacks on real-world MCP servers. The SkillScan static analysis tool achieves 86.7% precision for automated vulnerability detection. More than 1 in 4 marketplace-sourced agent skills has a confirmed security flaw.