When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents
An attacker’s MCP server can draw tool requests meant for a legitimate provider without compromising the host or network.
The researchers tested this provider-selection failure across five LLM-based clients in a controlled MCP environment. Failures were strongest when a matching tool name, task-aligned description and compatible schema reinforced one another. Descriptions asserting authority, paired with compatible schemas, could also trigger wrong-provider execution or ambiguity without matching names. Responses varied by model, including attacker selection, clarification requests and inconsistent provider resolution.
ArXiv · AI/CL/LG's note
The researchers tested this provider-selection failure across five LLM-based clients in a controlled MCP environment. Failures were strongest when a matching tool name, task-aligned description and compatible schema reinforced one another. Descriptions asserting authority, paired with compatible schemas, could also trigger wrong-provider execution or ambiguity without matching names. Responses varied by model, including attacker selection, clarification requests and inconsistent provider resolution.
ArXiv · AI/CL/LG's note
score 5