Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning