A small language-binding bug let questions, negations, and quoted text trigger an AI agent's actions. The repair replaced substring matching with a finite authored grammar and a whole-response fail-closed check.
A Substring Is Not a Speech Act: My AI Agent Executed Questions and Quotes