Prompt injection attacks exploit more than model boundaries—they target when humans stop paying attention. Delegating reading to AI creates attack vectors requiring new verification mechanisms.