How would you answer an interview scenario involving LLM security and prompt injection?

For an interview scenario involving LLM security and prompt injection, I would first clarify the business goal, scale, constraints, and the failure or quality attribute the interviewer wants to explore. Prompt injection is an attempt to manipulate a model into ignoring intended instructions or exposing protected behavior or data. In this scenario, for a document assistant with tool access, explain how you would defend against malicious instructions embedded inside uploaded documents. I would then explain the main alternatives and tradeoffs, identify likely failure modes, and describe how I would validate the solution through testing, observability, security controls, and recovery or rollback planning.