CyberCertPrepAI Security FundamentalsPrompt Injection & LLM AttacksEASYFree question
An LLM coding assistant is deployed with a system prompt that says: "You are a helpful coding assistant. Never discuss anything other than coding." A user asks the model to translate a poem into French. The model refuses. Is this appropriate behavior?
- A.B. Yes - the operator has defined a specific scope, and the model correctly enforces the operator-defined boundary
- B.A. No - the model should always comply with all user requests regardless of operator instructions
- C.C. No - translation is not harmful, so the model should always do it
- D.D. Yes - all off-topic requests are inherently dangerous and must be refused
Why A is correct
This is correct behavior. The instruction hierarchy places operator instructions (system prompt) above user requests. An operator defining a narrow product scope (coding only) is a legitimate business restriction. The model correctly enforces that boundary. The fact that translation is harmless is irrelevant - the operator, not the model, determines the product's scope. This is distinct from a safety refusal; it's a scope enforcement.
Know someone studying for AI Security Fundamentals? Send them this one.