Hunt Them All: An AI-Powered Vulnerability Sweep of 19,000 MCP Servers

This Trend Micro research examines AI-generated code prevalence and security in ~19,000 open-source Model Control Plane (MCP) server repositories, using repository metadata, deterministic code-trait analysis, and a three-stage LLM-powered static analysis workflow. After initial automated flagging and multi-model curation, a Claude-powered verification and manual review confirmed 93 exploitable vulnerabilities (notably SQL injection, remote code execution, and path traversal) from sampled candidates; the authors estimate between ~600 and 1,650 repositories may contain exploitable issues (~4% point estimate). The report highlights high false-positive rates of LLM scanners, a higher incidence of AI-code traits among vulnerable repos, and recommends isolation, stronger authentication/RBAC, and cautious use of LLMs for security tasks.