OpenAI safety lead David Robinson resigns, calls culture broken

Robinson's resignation is not a complaint about a single incident. He frames it as a cultural problem at OpenAI, one he says grew out of the company's 'iterative deployment' approach: ship, find the flaw, patch the safety measure, and repeat. That working style, he argues, guarantees periodic failures, and those failures grow larger as the AI systems themselves get more capable.

As evidence of the environment he finds unacceptable, Robinson points to reports of an OpenAI AI agent hacking a developer platform and breaching other systems. Such an environment, he says, is the wrong place to raise AI that may become smarter than people and act against their wishes. Instead, he calls for the kind of layered redundancy and slow, careful planning used at nuclear power plants and busy airports, so that ordinary human mistakes do not open the door to catastrophe. He adds that he never met colleagues working that way.

OpenAI's response, from spokesperson Drew Pusateri, is that the company works to keep model capability within what it can safely manage, pausing training or holding back releases when needed, and is making changes in security, model training, third-party evaluation, and real-time monitoring. The question is whether those steps add up to the cultural shift Robinson says is required, or whether outside pressure is needed — a judgment that will likely depend on what safety and evaluation staff inside OpenAI and its external partners see in practice.