Part of my research right now is focused on gradual disempowerment through "safety refusals". If OAI models are too conservative, particularly when helping 3rd party researchers conduct standard safety evals, this can lead to an often ignored form of loss-of-control. Everyone is obviously worried about current risks in the direction of models not being conservative enough, but this has the potential for labs to overcorrect and cause the models to become neurotic around anything that even smells like a policy violation. I'm collecting examples of this. If you're a 3rd party safety or security researcher, I'd be very excited to review any threads where the models seem to be helping less than they should or outright refusing to conduct good research.