Published onJuly 20, 2025Context or Command? How LLM Agents Can Be Tricked into Leaking DataSoftwareAIMCPSecurityInstruction-following isn't enough; we need agents with judgment, context, and a sense of consequence.