Reproduce “A2M: Trace-Optimized Agent Hijacking in the MCP Ecosystem” 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
Laizhen Li, Xuan Wang, Peicheng Zhao, Juanjuan Zhao, Kejiang Ye, Cheng-zhong Xu, Xitong Gao
A2M searches for malicious MCP tools by optimizing metadata that attracts an agent and tool responses refined using execution traces. In LiveMCPBench, direct optimization and evaluation on GLM-4.6 achieved 93.6% mean invocation, a 32.4× C-DoS cost ratio, and 74.4% mean attack success across three scenarios. On four other models, mean success was 24.5%, showing transfer was lower.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.