A black-box jailbreak method that treats jailbreak-prompt discovery as an evolutionary search in which a large language model itself acts as the crossover, mutation, and fitness operators over a population of jailbreak templates. Rather than gradient-based token optimization, it seeds a population from human-written templates, applies LLM-driven heuristic crossover and mutation guided by explicit objectives (attack success, stealthiness, conciseness), and selects survivors by attack-success fitness. It further frames transfer to the full harmful-query set as a transfer-learning step (evolve on a small clustered subset, then generalize), cutting query and time cost while achieving high, transferable attack-success rates across open- and closed-source models.
While safety-aligned large language models (LLMs) are increasingly used as the cornerstone for powerful systems such as multi-agent frameworks to solve complex real-world problems, they still suffer …