Обсуждение

i’m not gonna lie to you, of all the shit that’s gone down this week with openai, this is the thing that scares the absolute shit out of me

Clearly we haven't been leaning into the incentive constructively Benchmark models on their ability to find, and then _patch_ (rather than exploit) the vulnerability

It was a relief to me that the sandbox models were told to penetrate sites like Hugging Face. When the models start making their own goals, then I will start worrying.