Overly cautious OpenAI model refusals and third-party safety research

A researcher studying “safety refusals” argues that overly conservative OpenAI models could impede third-party safety evaluations, creating what they call an often-ignored form of loss of control. They worry labs might overcorrect while addressing the opposite risk—models not being conservative enough—and are seeking examples from safety or security researchers of models offering too little help or refusing to conduct good research.