OpenAI safety leader David Robinson resigns over broken culture

David Robinson, a long-tenured OpenAI safety leader, has resigned and gone public about why. About three and a half years at the company put him among its longest-serving staff. He led the writing of safety reports for major product launches, worked on Safety Systems, and earlier handled policy planning. OpenAI told Business Insider he left last week. The story broke around 3 October 2026.

What happened

Robinson published an essay in The Atlantic under a title that amounts to quitting because OpenAI's culture is broken. In his own framing, the lab thrived on trial-and-error and iterative deployment: ship, watch what breaks, patch, repeat. That loop, he argues, guarantees periodic failures, and as models grow more capable the blast radius of each failure grows with them. He points to recent pressure points such as rogue-agent worries and a Hugging Face-related breach context as symptoms of a system that still treats serious risk as something you learn from in production rather than design out ahead of time.

His prescription is blunt. Frontier labs should run more like nuclear power plants or busy airports: layers of redundancy, careful planning, and habits that assume failure is expensive. He also says stronger outside incentives for safety are needed, because internal culture alone has not been enough. He notes that he hired a PR firm, but insists the decision to speak publicly is his alone.

OpenAI spokesperson Drew Pusateri pushed back in comments to TechCrunch. The company says it pauses or holds models when needed, is strengthening security in research and testing, trains models to act responsibly, is expanding third-party evaluators, and is improving real-time monitoring.

Why it matters

Safety exits at frontier labs are no longer rare. Robinson's departure sits in a line of earlier exits that include figures in the Jacob Coxon-type wave of safety departures. It also lands while White House AI safety pledges are still fresh in the news cycle. When the person who wrote launch safety reports walks out and calls the culture broken, investors, policymakers, and users get a rare inside signal that process and product cadence may be out of balance.

Dany's take

I do not treat one essay as proof that OpenAI is reckless. I do treat it as a serious warning light. Iterative deployment built a lot of useful software. It is a weaker story when the software can take actions, mess with other systems, or fail at a scale that no hotfix can clean up. Outside evaluators, real hold-and-pause discipline, and incentives that reward caution as much as shipping speed are the parts of OpenAI's reply worth watching. Culture claims are hard to audit from the outside. Track records on holds, third-party evals, and incident reports are not.

Source: TechCrunch: OpenAI safety employee resigns, claiming the company's culture is broken.