AutoDAN-Turbo: the AI that writes its own jailbreaks so you don't have to
Ourtake
Nobody had to teach it to be sneaky. It worked out the sneaky part on its own.
What happened
A team of researchers from US universities and Nvidia created AutoDAN-Turbo, a system that automatically discovers and combines jailbreak strategies to get language models past their built-in rules. It needs only text output from the target model and reported an 88.5% attack success rate on GPT-4-1106-Turbo, rising to 93.4% with added human-designed strategies. The code is freely available on GitHub.