MCP tool poisoning in agentic AI frameworks: a cross-framework measurement study
AI summary
The arXiv paper by Farooq Shaikh and Biagio Montaruli (https://arxiv.org/html/2611.03555v2)对智能体 conducts a cross-framework measurement study of MCP tool poisoning in AI frameworks. The paper notes that the Model Context Protocol (MCP) is used to connect LLMs to external tool providers, but its design passes attacker-controllable content such as tool descriptions, schema fields, and response data directly into the model context without any validation.