Former OpenAI safety report author resigns, calling the industry's culture broken

David Robinson, who wrote the safety reports that accompanied OpenAI's major model releases, has resigned from the company and gone public with a critique of its safety culture. According to The Verge, which reported the departure on October 3, 2026, Robinson makes his case in an editorial published in The Atlantic, and his central verdict is a single word: "broken." He applies it both to his former employer and, more broadly, to the industry as a whole.

A critique of culture, not just rules

Per The Verge's account of the essay, Robinson argues the problem runs deeper than regulation or new procedures around model training. He describes the industry's culture as fundamentally broken, pointing to habits he says define Silicon Valley: extreme self-confidence, a permanent sprint pace, and an optimism that leads teams to ignore or underestimate potential problems while building ever-larger models.

His proposed fix is a change in posture rather than paperwork. Robinson says AI companies need to cultivate humility and deliberately seek expertise from outside the tech industry's insular, speed-first world. The analogy he reaches for is high-stakes infrastructure: frontier labs should operate like nuclear power plants or busy airports, with layered redundancy and slow, deliberate planning, so that an ordinary human error — inevitable at some point — does not open the door to catastrophe.

One departure in a growing exodus

The Verge situates Robinson's resignation within a steady stream of safety-focused staff leaving prominent AI firms. Jacob Coxon is described as having kicked off the pattern by quitting Anthropic and then warning publicly that AI might wipe out humanity within the decade. Per the report, he was followed out the door by Robert O'Callahan, Bilal Chughtai and Josh Engels at Google DeepMind, and Joe Benton at Anthropic.

The insider-warns-after-leaving problem

The Verge's author, weekend editor Terrence O'Brien, addresses the obvious reaction head-on: it is reasonable to feel cynical about people who helped build a risky technology and only sound the alarm once they have resigned. After all, they helped create the situation. O'Brien's counterpoint is that this cynicism, while understandable, is not a good reason to dismiss what they now say.

Why it matters

Robinson was not a peripheral figure at OpenAI; he authored the safety documentation attached to the company's flagship model releases. When the person who wrote those reports says the culture that produced them is broken, it raises a direct question about how much influence safety work actually has inside a frontier lab.

His argument also operates below the level of policy. Rules governing training practices can be written, audited and amended; a culture of unchecked confidence and permanent urgency is far harder to regulate from the outside, which is precisely why he frames it as the root problem rather than a symptom.

Finally, the pattern matters more than any single exit. With departures now recorded at OpenAI, Anthropic and Google DeepMind, the trend points to something industry-wide rather than one lab's management failure. For anyone tracking AI governance, a growing chorus of former safety staff choosing public warning over internal process is itself a data point on whether internal checks are holding as competition among the major labs intensifies.