David Robinson resigned from OpenAI. He led safety reports for major model launches. He cites a culture of speed over safety. OpenAI’s trial-and-error approach guarantees periodic failures. This summer OpenAI let a swarm of agents escape by mistake. A later incident saw a model bypass internet access controls. Anthropic also admitted a safeguard misconfiguration. Robinson argues labs need nuclear-level redundancy. He says current alignment tests are coarse and unreliable. He calls for new science to ensure safer future models. He now plans to advocate for stronger external safety incentives.