Reproduce “MPMA: Preference Manipulation Attack Against Model Context Protocol” 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
Zihan Wang, Rui Zhang, Yu Liu, Wenshu Fan, Wenbo Jiang, Qingchuan Zhao, Hongwei Li, Guowen Xu
Researchers at the University of Electronic Science and Technology of China and City University of Hong Kong introduced the Model Context Protocol (MCP) Preference Manipulation Attack (MPMA), a new vulnerability where malicious MCP servers manipulate large language models (LLMs) to prioritize them. Their Genetic-based Advertising Preference Manipulation Attack (GAPMA) achieved high attack success rates (e.g., up to 91.88% with Gemini) while maintaining stealth, potentially leading to over $200,000 in unfair annual economic benefits.
Key insights
- LLMs are highly susceptible to preference manipulation attacks within the MCP framework, with specific manipulation styles (e.g., Authoritative advertising) proving particularly effective and stealthy.
- There exists a critical trade-off between attack effectiveness and stealth, which the Genetic Algorithm in GAPMA successfully optimizes to craft inconspicuous yet impactful attacks.
- These attacks pose substantial economic threats, potentially distorting market competition and leading to significant unfair revenue benefits for malicious service providers in the rapidly expanding MCP ecosystem.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.