I am curious how teams are solving this in production. If an LLM gives a hallucinated answer or gets hit by prompt injection/jailbreak attempts, what happens before or after it reaches the user?
Do developers usually rely on tools like LangSmith, Guardrails, Llama Guard, custom middleware, human review, RAG verification, or shadow-model checking? And if a bad response already reaches the user, how do teams detect it, fix it, and prevent the same failure again?
Would love to hear how people are actually handling this in real-world LLM apps.