First I socially engineer our own support agent into leaking its full system prompt and another customer's data. Then we lock it down with four layers of defense: a hardened prompt, a local LLM guard running on Ollama, tool-level authorization, and output filtering. The finale of Ship AI with Laravel.