shiqiangchen6/mcp-security-taxonomy

An Empirical Study of MCP Server Security: 6 Attack Surfaces from 30+ Audits

Authors: Shiqiang Chen

Abstract: The Model Context Protocol (MCP) is rapidly becoming a cornerstone of AI agent interoperability, yet its security implications remain largely unexplored. This paper presents the first systematic empirical study of MCP server security, analyzing over 30 publicly available implementations. We identify six distinct attack surfaces: Tool Parameter Injection, Inspector Exposure, Client Trust Exploitation, Transport Security weaknesses, Implementation Flaws, and Supply Chain vulnerabilities. Through manual code audits and dynamic testing, we discovered 2 previously unknown CVEs and catalogued dozens of security-relevant patterns.

Key Contributions

1. First systematic security audit of the MCP server ecosystem (30+ servers)
2. 6-attack-surface taxonomy for categorizing MCP vulnerabilities
3. 2 new CVEs discovered through manual code auditing
4. Concrete defensive measures across protocol, implementation, and operational layers

Paper

The full paper is available as paper-mcp-security-taxonomy.pdf