How should LLM security and prompt injection be designed for production?

Keep authorization and secrets outside model control, treat retrieved text as untrusted data, restrict tools, validate outputs, minimize sensitive context, and test adversarial inputs.