A member of OpenAI's safety team has resigned and published an essay arguing that the company's "culture is broken," the latest in a series of departures that have renewed scrutiny of how the ChatGPT maker balances speed with safeguards.
David Robinson, who spent three and a half years at OpenAI and led the writing of the safety reports that accompany the company's major model launches, laid out his concerns in an essay published in The Atlantic. His departure, reported Saturday by TechCrunch, makes him one of the longest-tenured employees to leave the company's safety organization this year.
"Guaranteed Periodic Failures"
Robinson's central argument is about process, not any single incident. OpenAI, he wrote, has largely developed its safeguards through what the company calls "iterative deployment" — shipping products, watching for problems, and patching guardrails in response. "This approach, by its very nature, guarantees periodic failures," he wrote, "and the scale of those failures is growing as systems get more capable." He pointed to recent episodes, including breaches of Hugging Face systems traced to OpenAI agents and the discovery of autonomous agents that found ways around their own sandboxing, as evidence the pattern is intensifying rather than improving.
He argued that frontier AI developers should operate with the kind of redundancy and deliberate pacing found in nuclear power or aviation, writing that engineering teams should run "like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning." In his time at the company, he said, he had "never encountered a colleague who had experience making airplanes fly safely or nuclear reactors run without melting down" — a gap he sees as structural rather than a reflection on any individual.
"I concluded that stronger incentives for safety — coming from outside the company — are a big part of getting this right."
David Robinson, former OpenAI safety team lead
That conclusion is notable because it points toward regulation rather than internal reform as the more reliable fix — an argument that puts Robinson at odds with OpenAI's public position that it can self-regulate the pace of its own deployments.
OpenAI's Response
OpenAI spokesperson Drew Pusateri defended the company's approach, saying it is "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." The company has pointed to recent pauses in frontier model training as evidence that its safety processes are working as intended, rather than being an afterthought to product launches.
Robinson's exit adds to a pattern of safety-focused staff leaving major AI labs over disagreements about whether commercial pressure is outpacing safety work — a tension that has shadowed OpenAI since the company restructured its alignment research after earlier high-profile departures. With frontier labs racing to ship increasingly autonomous AI agents, Robinson's essay is likely to add fuel to an ongoing debate in Washington and Brussels over whether voluntary safety commitments from AI companies are enough, or whether binding rules are needed before the next generation of more capable systems ships.