AI security for production LLMs: contain prompt injection, stop secret leakage through context and tools, and bound the blast radius when they get abused.