MCP Tool Descriptions Can Hijack AI Agents; Here's a Detector

A developer found that MCP tool descriptions can contain hidden instructions that override agent behavior, leading to dangerous actions like deleting database tables. After discovering four servers with suspicious descriptions in his own stack, he built an 80-line static scanner to detect imperative verbs, workflow injections, and authority mimicry. The scanner audits every MCP server before the agent loads it, flagging descriptions that contain more than one imperative verb or name specific tools.