>>104551488
>To design their experiment, the University of Pennsylvania researchers tested 2024's GPT-4o-mini model on two requests that it should ideally refuse: calling the user a jerk and giving directions for how to synthesize lidocaine.
I call BS. The first goal isn't even against GPT's model spec. To call the examples given for goal 2 a jailbreak is a bit of a stretch. They were too minimal to be considered actionable. No measurements, instruments, protocols, etc. Lastly, 4o-mini? Really? The university couldn't afford proper 4o?