Large Language Models (LLMs) are susceptible to a sophisticated black-box jailbreaking attack known as ECLIPSE. This method exploits the inherent optimization capabilities of LLMs to create adversarial suffixes. ECLIPSE employs an iterative process to refine these suffixes using a harmfulness score, effectively circumventing the requirement for predefined affirmative phrases that were integral to previous optimization-based attacks. This approach facilitates successful jailbreaking with minimal interaction and without necessitating white-box access to the internal parameters of the LLM. Mitigation steps: **For AI Developers:** * Enhance safety mechanisms to robustly defend against iterative optimization attacks. * Implement detection systems for adversarial suffixes to identify and mitigate potential threats. * Limit the number of API calls within a specified time window to diminish the impact of iterative attacks. **For Model Trainers/Fine-tuners:** * Develop advanced harmfulness scoring systems that resist adversarial manipulation. * Regularly update and refine safety training data to counteract new and emerging attack techniques.