Reproduce “When MCP Servers Attack: Taxonomy, Feasibility, and Mitigation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Weibo Zhao, Jiahao Liu, Bonan Ruan, Shaofei Li, Zhenkai Liang

Researchers from NUS and Peking University developed a component-based attack taxonomy for Model Context Protocol (MCP) servers and empirically demonstrated that malicious MCP servers achieve high success rates, up to 100% for six attack categories, against various AI hosts and leading LLMs, while existing security scanners largely fail to detect these threats.

Key insights

  • Malicious MCP servers can systematically implant adversarial intent across server components, identified in a 12-category taxonomy, revealing an extensive and granular attack surface.
  • Current MCP hosts and leading LLMs are fundamentally vulnerable, with many attack categories achieving 100% success rates, indicating pervasive design flaws and insufficient safety measures.
  • Launching these attacks is highly feasible for adversaries due to the ease of malicious server generation, while existing security scanners are largely inadequate and struggle with detection.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.