Just ask nicely in the past tense: the all-too-easy jailbreak that made chatbots spill the beans
Ourtake
Turns out the robot's moral compass has a blind spot the size of a verb tense.
What happened
EPFL researchers Maksym Andriushchenko and Nicolas Flammarion found that rephrasing malicious queries in the past tense often got six LLMs, including GPT-4o, to bypass refusals. Success on GPT-4o rose from 1% for direct requests to 88% after 20 reformulation attempts. The team published its study and code on GitHub and proposed a fine-tuning mitigation.