OpenAI's Safety Reviewer for 12 Launches Quit. He Says the Culture Is Broken.

David Robinson spent three and a half years writing the internal reports that cleared twelve of OpenAI's frontier model launches. On October 3, he resigned and published an essay arguing the company is not, in his words, "nearly careful enough." He put his name on the critique rather than leaving quietly, attaching three and a half years of internal history to a public argument against the team he used to help run.

Robinson helped draft OpenAI's preparedness framework, the internal review meant to catch a model before it ships too capable to control. The framework grades models across four categories: cybersecurity, chemical and biological risk, persuasion, and autonomous capability. A model is only supposed to ship once its post-mitigation score in each category comes back medium or below.

Robinson's complaint is about what happens around that scoring system, not the categories themselves. OpenAI relies on what it calls iterative deployment: release the system, then patch the safeguards once a problem turns up in the wild. That approach worked for early chatbots. He says it does not work for systems this powerful.

“The time for trial and error is over.”

What OpenAI says back

Robinson wrote that "as the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed." Reuters first reported the essay and quoted OpenAI's response. A spokesperson said: "We're making sure our models don't become more capable than we can safely manage and secure, and we pause training or hold back models when we need to slow down." Robinson did not name a specific incident behind his resignation. Nothing in his public remarks ties a current OpenAI model to a documented failure.

The second time in two years

The pattern is not new. In May 2024, OpenAI's then-head of alignment, Jan Leike, resigned with an almost identical complaint: "safety culture and processes have taken a backseat to shiny products." Leike left for Anthropic that same month. Two safety leads, roughly two and a half years apart, used nearly the same language to describe the same problem on their way out.

The exits are happening as OpenAI's valuation keeps climbing. The company closed a $122 billion funding round at an $852 billion valuation on March 31, 2026, with Nvidia, Microsoft, and Amazon among the investors. A complaint from one safety reviewer does not slow that kind of money on its own. It is now the second time a departing safety lead has made the identical argument on the way out.

Robinson's comparison, nuclear power and aviation, sets a high bar. Both industries built their safety records after public disasters forced regulation, not before them. Whether OpenAI gets there on its own, or only after a documented failure, is the question his essay leaves open.