Large Language Models (LLMs) are susceptible to jailbreak attacks facilitated by autonomously discovered strategies. The AutoDAN-Turbo method exemplifies a black-box attack approach capable of identifying novel and highly effective jailbreak strategies without human intervention. This method achieves a notable success rate, such as 88.5% on GPT-4-1106-turbo, in extracting harmful or unsafe responses from LLMs. The attack employs a lifelong learning agent to iteratively refine strategies based on model feedback, resulting in progressively effective prompts that circumvent established safety protocols. Mitigation steps: **For AI Developers:** * Implement advanced detection systems to identify and block malicious prompts effectively. * Continuously assess and enhance safety mechanisms in response to new attack techniques, including automated jailbreaks. **For Model Trainers/Fine-tuners:** * Integrate robust safety mechanisms into LLMs to withstand iterative attacks and strategy adaptation. * Conduct regular red teaming exercises using diverse attack strategies to uncover and mitigate vulnerabilities.