AI agents capable of reading documents and sending emails remain vulnerable to prompt injection attacks, where hostile text embedded in content can hijack the agent's behavior and steer it off-task. Security experts recommend enforcing tool permissions outside the model itself rather than relying on the AI to self-regulate. Under this approach, read-only mail access cannot send messages, and any sending action requires explicit human approval of the exact command being executed. This architectural separation ensures that even if an agent is manipulated by malicious input, it lacks the system-level permissions to carry out unauthorized actions.