Large Language Models (LLMs) are revolutionizing how we interact with technology, but they're not without their weaknesses. One such vulnerability is their susceptibility to "jailbreak attacks," carefully crafted prompts designed to trick LLMs into generating harmful or inappropriate content, bypassing their safety protocols. Existing methods for crafting these jailbreak prompts have limitations. Manual methods require significant effort and lack consistency, while automated optimization techniques often produce unnatural, easily detectable prompts or demand high computational resources. Enter GASP (Generative Adversarial Suffix Prompter), a new framework that efficiently generates adversarial suffixes—short phrases added to the end of a prompt—to jailbreak LLMs while maintaining a natural, human-readable style. Unlike previous attacks, GASP operates in a black-box setting, meaning it doesn't need access to the LLM's internal workings. It achieves this by leveraging Latent Bayesian Optimization (LBO) to explore the vast space of possible suffixes within a continuous embedding space. This approach avoids the computational bottleneck of traditional discrete token optimization. GASP further refines its attack strategy using a custom evaluator, GASPEval, and Odds Ratio Preference Optimization (ORPO). GASPEval assesses the effectiveness of generated suffixes, providing feedback to continuously improve the attack. ORPO fine-tunes the model, prioritizing high-success suffixes while preserving readability. Experiments show that GASP significantly outperforms existing methods, achieving higher attack success rates across various open-source and proprietary LLMs. Its efficiency is a standout feature, boasting faster training and inference times. Furthermore, human evaluations confirm that GASP-generated prompts are more readable and less suspicious than those created by other methods. GASP represents a significant advancement in understanding and exploiting LLM vulnerabilities. By shedding light on these weaknesses, GASP contributes to the ongoing development of more robust LLM defenses, paving the way for safer and more reliable AI systems. While GASP can be used to expose vulnerabilities, it's important to remember that manual jailbreak methods already exist. The researchers behind GASP emphasize responsible disclosure, sharing their findings with relevant organizations to improve LLM security.