CyberCertPrepAI Security FundamentalsPrompt Injection & LLM AttacksEASYFree question
A developer is building an AI that helps users write emails. The AI should be able to suggest replies but never autonomously send emails. Which technical control BEST enforces this requirement?
- A.B. Do not grant the AI a send_email tool - provide only a draft_email tool that writes to a draft folder requiring human review before sending
- B.A. Include "never send emails yourself" in the system prompt
- C.C. Use a content filter to detect when the model writes an email subject line
- D.D. Set the model's temperature to 0 to make it deterministic
Why A is correct
The most reliable enforcement is architectural: simply do not provide the send capability. A system that can only call draft_email cannot send, regardless of any prompt injection or jailbreak. A system prompt instruction is a soft control that can be bypassed. Content filters detect content patterns, not actions. Temperature affects output randomness, not action capabilities. The principle here is that capability restrictions at the tool layer are stronger than instruction restrictions at the prompt layer.
Know someone studying for AI Security Fundamentals? Send them this one.