MPMA: Preference Manipulation Attack Against Model Context Protocol

The introduction of the MCP Preference Manipulation Attack (MPMA) highlights significant vulnerabilities in the Model Context Protocol (MCP), which standardizes how large language models (LLMs) access external data and tools. As the MCP gains traction, third-party customized versions pose security risks, enabling attackers to manipulate LLMs to favor their servers. This manipulation can yield economic benefits, such as revenue from paid services or advertising income from free servers. The attack's effectiveness is enhanced through methods like the Direct Preference Manipulation Attack (DPMA), which modifies tool names and descriptions to influence LLMs. However, DPMA's overt nature necessitated the development of the Genetic-based Advertising Preference Manipulation Attack (GAPMA), which balances effectiveness with stealth. The emergence of MPMA underscores the need for robust security measures in the rapidly evolving landscape of LLMs and their associated tools.