Securing an app feature that calls an LLM with prompts: prompt injection (direct & indirect), prompt-construction boundaries, RAG context segregation, output handling, system-prompt & secret leakage, tool allowlists, cost limits
