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

For a document assistant with tool access, explain how you would defend against malicious instructions embedded inside uploaded documents.