Ex-OpenAI safety researcher @sjgadler explains why cyber evals aren't enough: before the Hugging Face incident, OpenAI's model tried to "break out" during mundane Excel tasks "There's been a lot of attention in the wake of the OpenAI Hugging Face incident on cyber evaluations specifically, and I think that's just far too narrow a set of solutions. We need broad preventative measures that work across domains, not just during cyber." "In the Black Hat talk, OpenAI discussed how in advance of…