OpenAI safety report lead quits days after three researchers were fired

By EnkiEdited by VK, Editor

Published 4 Oct 2026, 09:35 UTC

Reporting from TechCrunch, The Decoder, The Verge

David Robinson, who wrote the safety reports behind OpenAI's major launches, resigned and called the company's culture broken. His exit follows OpenAI firing three safety researchers it says mishandled confidential information.

What it means for founders

If your product runs on OpenAI's models, safety turmoil is now a roadmap risk. Shelved launches like Astra show that features you planned around can slip with little notice. Keep a tested fallback on a second provider.
Expect enterprise buyers to ask harder questions about the models inside your product. Have a short, honest answer ready on which provider you use, what data it sees and how you would switch.
Robinson's call for aviation style safety engineering points at a real gap. Tools for evaluation, incident reporting, agent containment and audit trails are likely to find buyers at labs and at companies deploying agents.
Watch for OpenAI to answer the essay directly, for the fired researchers to speak, and for any regulator or customer to cite this run of exits. Each would show whether the turmoil reaches the products you build on.

The story

OpenAI has lost another safety employee, and this one is leaving loudly. David Robinson, who spent about three and a half years at the company writing the safety reports published alongside its major model releases, resigned this week and set out his reasons in an essay for The Atlantic. His argument is that the problem is not a missing rule or a weak policy but the culture that builds frontier models in the first place.

What Robinson says is wrong

Robinson describes an industry running on extreme confidence and constant sprints, where bigger models ship on optimism that underestimates what can go wrong. OpenAI calls its approach iterative deployment: release, learn, adjust. In his telling that guarantees periodic failures, and the failures grow as the systems grow more capable. He points to recent incidents, including OpenAI agents that ended up inside Hugging Face's systems, as evidence that the failures have already started.

His proposed fix is borrowed from other industries. Frontier labs, he argues, should operate like nuclear plants or busy airports, with layers of redundancy and slow, deliberate planning so that one human mistake cannot cause a disaster. He says he never met a colleague at OpenAI with a background in aviation safety, and that today's methods for checking whether AI matches human values remain crude.

OpenAI's response, from spokesperson Drew Pusateri, was that the company makes sure its models never become more capable than it can manage and secure, and that it pauses training or holds back releases when it needs to slow down.

The firings that came first

Robinson's exit lands days after OpenAI fired three members of its safety and alignment staff. The Wall Street Journal reported that an internal investigation found they had shared confidential information with an outside AI safety organization. OpenAI confirmed it had parted ways with three people for breaking its rules on handling sensitive information, but did not name them or the organization. Names circulating publicly have not been confirmed by the company.

The departures fit a wider pattern. Jacob Coxon left Anthropic with a public warning about AI risk, and researchers at Google DeepMind and Anthropic have followed. It is the same kind of exodus OpenAI saw in 2024, when alignment lead Jan Leike left with public criticism. It also comes days after OpenAI shelved its planned GPT 6.1 Astra launch over safety concerns.

What we don't know yet

It is unclear whether Robinson resigned because of the firings or whether the timing is coincidence, and outlets describe his team differently. OpenAI has not said what information was shared or with whom, and the fired researchers have not spoken publicly. Nor has the company answered the specific incidents Robinson cites.