For months, David Robinson's job at OpenAI was to write the safety reports released alongside each new model, the documents meant to reassure the public that someone, somewhere, had checked for danger before the technology shipped. This week, he quit, and in an editorial for The Atlantic, he said the culture he left behind is not simply flawed but "broken" at its foundation. It would be easy to dismiss this as the latest case of an insider discovering a conscience only after the product was already built and sold, from San Francisco to Singapore. Robinson acknowledges as much himself.

Yet his argument is more specific than regret: he says Silicon Valley's habit of "perpetual sprints" and "extreme confidence" has produced an industry that treats caution as friction rather than necessity. What is needed, he argues, is not another patch of rules but a different temperament altogether, one borrowed from industries that already understand how catastrophic a single mistake can be. "Given today's risks, frontier labs need to run like nuclear power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster," Robinson writes, comparing the stakes of frontier AI development to systems where failure is measured in lives, not headlines. He is not alone.

Over recent months, at least five researchers have left major labs voicing similar alarm: Jacob Coxon departed Anthropic warning that AI "could kill us all by the end of the decade," followed by Robert O'Callahan, Bilal Chughtai, and Josh Engels at Google DeepMind in London and Mountain View, and Joe Benton at Anthropic. Taken individually, each resignation might be dismissed as one dissatisfied employee.