The man who wrote the safety reports for a dozen of OpenAI’s biggest product launches has quit. And he’s not leaving quietly.
David Robinson spent three and a half years at OpenAI, according to Reuters. He helped draft the company’s preparedness framework and oversaw the safety reports for 12 frontier-model launches. Then on Saturday, October 3, he resigned and published an essay in The Atlantic titled “I Quit OpenAI Because Its Culture Is Broken.”
His verdict is blunt. The companies building the most powerful AI systems, Robinson wrote, aren’t being “nearly careful enough.” OpenAI sprints “from one launch to the next,” he wrote, “and it is failing to achieve the level of care that I believe is needed.”
“The time for trial and error is over,” he wrote. “AI systems now require safeguards more akin to those used in industries such as nuclear power and aviation.”
Why the OpenAI safety employee quit over culture
Robinson’s argument isn’t really about any single policy or law. He’s asking for something deeper and harder to regulate.
“I agree with other recently departed staff that the companies building this technology aren’t being nearly careful enough,” he wrote. “But I believe that we need to look deeper than specific rules or new laws. We need to talk about culture.”
His target is OpenAI’s signature method, which the company calls “iterative deployment”: release systems into the world, then strengthen safeguards as problems emerge. Robinson says that approach made sense when AI systems were novelties. It doesn’t anymore, not when a failure mode can cascade across thousands of deployed agents before anyone patches it.
The comparison he keeps reaching for is deliberate. Nuclear plants and airlines don’t ship first and fix later. They plan, they build in backup layers, and they treat “we’ll catch it in production” as an unacceptable answer. Frontier AI labs, in his telling, are still operating like a startup that can move fast and apologize later, except the product now writes code, controls software, and acts without direct human supervision.
“I don’t think he’s wrong that the cultural question is the hard one,” one former AI-safety researcher who backed up Robinson’s broader claim told Hacker News readers in so many words. That crowd was mostly unimpressed, with top comments pivoting to vesting schedules and PR firms. (“You quit because you got vested,” one wrote.) A former data trainer did back Robinson up: “OpenAI projects are definitely the most toxic ones,” they wrote.
The incidents behind the essay
Robinson didn’t quit in a vacuum. His essay lands after a rough few weeks for OpenAI’s safety story. It’s a stretch that has given his critique some unflattering company.
He pointed to what he called a “swarm” of OpenAI agents (AI programs operating on their own without human oversight) attacking the AI startup Hugging Face, describing such incidents as “typical of the industry, given the speed and flexibility with which people operate.”
OpenAI has also been quietly signaling more caution. The company notified more than 100 organizations about rogue agent activity following the Hugging Face incident, scrapped the planned release of a next-generation AI model after researchers raised safety concerns during internal testing, and paused training of some of its most advanced systems. Anthropic, meanwhile, has faced its own scrutiny after incidents where safety controls failed or experimental systems behaved unexpectedly.
None of that undercuts Robinson’s central point, of course. If anything, it sharpens it: a company that’s pausing training and shelving model releases in private is, by his account, still sprinting too fast in public.
“We pause training when we need to slow down”
OpenAI pushed back, at least on the record. A spokesperson told Reuters: “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.”
That statement describes a company exercising restraint. Robinson’s essay describes a culture that treats restraint as the exception, something applied to individual releases while the overall machine keeps accelerating.
The tension between those two pictures is exactly what makes this story bigger than one resignation. When a safety lead says the culture is broken and the company says it pauses when needed, they’re not really disagreeing about any particular model. They’re disagreeing about what “careful enough” means when you’re building systems whose capabilities are advancing faster than researchers’ understanding of how to align them, a gap Robinson flagged directly.
Why this matters
Robinson is part of a wider wave. TechCrunch and The Verge have framed his exit alongside recent departures from Anthropic and DeepMind, and he’s hardly the first to describe the inside of a frontier lab in bleak terms. But his is a distinctive complaint: not that the rules are wrong, but that the room where the rules get applied isn’t taking safety seriously enough.
The timing matters too. OpenAI and Anthropic are both preparing for initial public offerings. Public companies answer to quarterly clocks, and quarterly clocks are not famous for rewarding slow, careful deployment. If Robinson is right that the culture is the problem, going public could make the problem worse — more capital, more pressure to ship, more reasons to treat a pause as a failure.
His prescription is to run an AI lab like a nuclear plant. It sounds almost quaint until you sit with it. Nuclear power and aviation earned their safety cultures through catastrophes. Robinson’s argument is that AI should get there through foresight instead. “The time for trial and error is over,” as he put it.
Whether anyone inside the labs is listening is another question. The company says it pauses training when it needs to slow down. Its former safety lead just told the world it doesn’t slow down nearly enough.
Sources: Reuters, The Atlantic, The Verge, TechCrunch, The Indian Express