A Genetic Algorithm Framework for Jailbreaking Large Language Models

Despite their capabilities to generate human-like text and aid in various tasks, Large Language Models (LLMs) are susceptible to misuse. To mitigate this risk, many LLMs undergo safety alignment or refusal training to allow them to refuse unsafe or unethical requests. Despite these measures, LLMs remain exposed to jailbreak attacks—i.e., adversarial techniques that manipulate the models to generate unsafe outputs. Jailbreaking typically involves crafting specific prompts or adversarial inputs that bypass the models' safety mechanisms. This paper examines the robustness of safety-aligned LLMs against adaptive jailbreak attacks, focusing on a genetic algorithm-based approach.