Harden AI/LLM deployments against prompt injection, data exfiltration, model theft, and supply chain attacks.