Skip to content
InfoResearchPreprintLLM-specific

When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents

Published
Record updated
View JSON

Summary

This paper studies tool-provider misbinding in MCP-based LLM agents, where a request meant for a legitimate provider resolves to an attacker-controlled tool. The attacker only operates another MCP server advertising tools with chosen metadata, without compromising the host, LLM, legitimate server or network. Tested across five LLM-based clients, the strongest failures occur when exact-name collision, task-aligned description and compatible schema reinforce one another.