Formal Security Framework for MCP-Based AI Agents: Threat Taxonomy and Defense Mechanisms

arXiv·high signal

Researchers published the first formal security framework for Model Context Protocol (MCP) agents, now governed by the Linux Foundation's Agentic AI Foundation. The paper provides a structured threat taxonomy, verification models, and concrete defense mechanisms for the protocol that has become the de facto standard for connecting LLM agents to external tools. Directly actionable for anyone building or deploying MCP servers.