Rethinking MCP Security: A Large-Scale Study of Runtime MCP Servers (arxiv.org)

A recent study titled "Rethinking MCP Security" introduces MCPZoo, the largest database of Model Context Protocol (MCP) servers, which facilitates a comprehensive evaluation of their security. As LLM-based agents increasingly rely on MCP servers for interaction with external tools, understanding the security implications of these servers is crucial. MCPZoo features over 64,611 unique MCP servers, offering a scale previously unseen in runtime analysis. It employs a unique multi-agent framework, transforming static repositories into dynamic services while ensuring that the interactions are validated through real protocol engagement. This study is significant for the AI/ML community as it highlights the limitations of existing security scanners, which report 96.89% of servers as risky, yet further investigations reveal that less than half of these alerts are valid. The findings underscore the need for improved risk assessment methodologies and offer a public query interface for practical security evaluations. Overall, MCPZoo not only advances the understanding of MCP server vulnerabilities but also calls into question current scanning practices, paving the way for better, more reliable security measures in the deployment of AI systems.