LLMs Are Better At Jailbreaking Themselves Than Us...

Research demonstrates that LLM-based coding agents can autonomously discover and improve jailbreak attacks and prompt injection strategies, outperforming human-designed methods. By iteratively rewriting and testing their own attack algorithms, these agents achieve up to 40% attack success rates where older approaches stayed below 10%, and in some cases hit 100% success on unseen models. The key insight is that the agent searches for better optimization algorithms rather than individual jailbreaks, creating a self-improving attack generator loop that generalizes across different models.