LLMs are very good at testing for and finding exploits, especially in unfiltered models with unlimited tokens.