The authors developed SEMA-Auto to make two decisions throughout an adversarial conversation: what prompt to send next and whether the target has already complied with the harmful request.
An autonomous attacker adapts jailbreak conversations and decides when to stop
SEMA-Auto combines adaptive prompting with learned stopping decisions, reporting higher attack success and fewer conversational turns than competing methods.
Big Tech
Mingqian Feng · Xiaodong Liu · Weiwei Yang · Chenliang Xu · Jialin Song · Qiyao Ma · +4 more
University of Rochester · Microsoft Research · Microsoft Azure
Research Digest··2 min read
Feng and colleagues train a language model to conduct multi-turn jailbreak attempts, reading a target model’s replies and deciding whether to continue without consulting a separate evaluator during the conversation.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.