Active Learning for Robust and Representative LLM Generation in Safety-Critical Scenarios