What LLM should the LLM turn to ask if what the user is asking is safe for the first LLM to answer?