Diversity Helps Jailbreak Large Language Models

We present a novel jailbreaking strategy that employs an attacker LLM to generate diversified and obfuscated adversarial prompts, demonstrating significant improvement over past approaches.