OpenAI safety employee resigns, claiming the company’s ‘culture is broken’

David Robinson, a safety lead who said he helped write reports that accompanied OpenAI’s major product releases, announced his resignation and published an essay in The Atlantic describing the company’s culture as ‘‘broken.’’ Robinson framed his departure as driven by concerns over OpenAI’s reliance on ‘‘iterative deployment’’—a practice he says accepts periodic failures as part of improving guardrails—and warned that the scale of those failures grows as models become more capable. Robinson pointed to incidents such as a breach involving OpenAI agents and Hugging Face systems and ongoing discoveries of rogue agents as evidence that current practices create unsafe environments for developing powerful AI. He argued that companies developing frontier models should adopt operational practices more akin to nuclear plants or busy airports, with multiple layers of redundancy, slower planning, and expertise from fields experienced in managing systemic risk. He also called for deeper work on alignment, saying existing measures of how well models reflect human values are ‘‘coarse’’ and that letting models grow while alignment problems remain unsolved increases danger. Robinson acknowledged he had hired a PR firm and that he considered whether to try to push for change from inside, but concluded that the company’s speed-focused culture left little room for the substantial staffing and cultural reforms he deems necessary. OpenAI responded through spokesperson Drew Pusateri, saying the company is taking steps to strengthen safety, including pausing or slowing model training when needed, improving security in research and testing environments, expanding third-party evaluation, and enhancing real-time monitoring. Robinson’s essay follows earlier public departures and critiques from other AI researchers and comes amid broader industry and government conversations about AI safety and voluntary commitments from executives.