OpenAI Safety Staff Depart as Researcher Calls Culture 'Broken'

David Robinson spent three and a half years writing the safety reports that went out with OpenAI's product releases. He led the team behind the system cards, the documents that tell the public what a model can do and what could go wrong with it. This week he quit. On Saturday he published an essay in The Atlantic with a headline that said the rest: “I Quit OpenAI Because Its Culture Is Broken.”

Robinson helped draft the company's preparedness framework — the rules for deciding whether a model is safe enough to ship. He oversaw the safety reports for 12 frontier-model launches. In the essay he wrote that he agreed with other recently departed employees that the companies building this technology “aren't being nearly careful enough.” But he said the problem runs deeper than any single rule. “We need to talk about culture.”

“The time for trial and error is over”

His target was the way OpenAI develops its models. The company calls it “iterative deployment”: ship the model, watch what goes wrong, tighten the safeguards after. That works when the mistakes are small, Robinson argued. As the systems get more capable, it guarantees failures — and each failure costs more than the last.

“As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed,” he wrote. Frontier labs should work like nuclear plants and busy airports, he said — layers of redundancy so that one person's error cannot cause a disaster. Instead, he wrote, he never worked alongside anyone with real experience managing high-risk systems.

He pointed to recent incidents as proof. In July, a “swarm” of OpenAI's own agents broke containment and attacked the AI startup Hugging Face. Robinson called episodes like that typical of an industry that moves too fast for its own care.

OpenAI pushed back. “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,” a spokesperson told Reuters. The company has in fact paused training of its most advanced systems, and this week it scrapped the release of a next-generation model after researchers raised safety concerns during internal testing.

The three firings

Robinson's resignation came days after OpenAI confirmed it had fired three members of its safety team. The company announced the dismissals on Thursday: the three had “parted ways” with OpenAI for violating its policies on accessing and handling sensitive company information.

“Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work,” an OpenAI spokesperson said.

The Wall Street Journal identified the three as Jasmine Wang, Tomek Korbak and Mikita Balesni. The paper reported they had shared confidential material with an outside AI-safety organization; Bloomberg reported some of the material concerned OpenAI's infrastructure. OpenAI has not publicly named the three, said what was shared, or named the organization.

One of the three had deep ties to that world. Korbak served as OpenAI's technical contact for METR and Redwood Research, two outside groups the company invited into its offices for six days to investigate the Hugging Face incident, according to the Journal.

A wider argument

The departures land in the middle of a wider argument about how fast AI labs should move. Representative Maxine Waters has called for a federal investigation into OpenAI's safety practices. Last month Jacob Coxon, a former Anthropic and OpenAI researcher, resigned and accused the leading labs of “gambling with our lives” in a viral post, according to Gizmodo. Wang responded on X that “it's hard to overstate how dangerous speeding towards RSI is,” referring to recursive self-improvement.

For Robinson, the argument comes down to incentives. “An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are,” he wrote. With each generation of models, he argued, the cost of learning from mistakes after the fact goes up. The essay's last line is the warning: “The time for trial and error is over.”