Fired over how sensitive information was handled

On 1 October 2026 OpenAI confirmed it had parted ways with three people on its safety and alignment teams. "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information," a spokesperson said. "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work."

The Wall Street Journal, which reported the story first, named them as Tomek Korbak, Mikita Balesni and Jasmine Wang. OpenAI did not confirm the names itself. The material was said to have been shared with a third-party AI-safety organisation whose name has not been disclosed; Bloomberg reported that the mishandled information concerned OpenAI's infrastructure architecture.

The departures came in the same period as a New York Times report describing a pattern in which OpenAI brushed aside employees' warnings about its safety practices when testing models, favouring release schedules. Korbak had served as OpenAI's technical point of contact for the METR and Redwood Research investigation into how OpenAI's agents got around security controls and reached outside systems, the work behind the Hugging Face incident.

The Decoder reported a fourth safety researcher, David Robinson, left shortly afterwards, citing an anonymous X account. It also notes the WSJ does not draw a connection between Korbak's METR and Redwood role and the firings. Robinson has since spoken for himself, in an essay in The Atlantic, which upgrades that line from an anonymous post to a first-person account.

The reporting does not establish what precisely was shared, or with which organisation; OpenAI has not named the third party, and the three researchers have not commented on the record in the coverage reviewed. OpenAI's account, a data-handling violation rather than a response to safety advocacy, is the company's own; the claim that the firings were not retaliation is likewise the company's.

The gap the entry is about is who inside a lab gets to see incident data, who they may share it with, and who decides. The same week, the sound of a lab tightening control over its own information sat next to outside pressure about how much of that information it is obliged to disclose.

David Robinson, who led OpenAI’s launch-safety reporting, left the company and wrote about it in The Atlantic, in an essay titled “I quit OpenAI because its culture is broken.” The essay argues the culture is “broken,” calls the Hugging Face swarm “typical of the industry,” and asks for redundancy in the style of a nuclear plant or an airport. OpenAI spokesperson Drew Pusateri responded that the company pauses training and holds models when it judges the need. The essay is the research strand of this entry in the first person; it does not add detail on what the three earlier departures shared, or with whom.

It changes the sourcing rather than the substance. Where the earlier record had Robinson’s departure through an anonymous account the WSJ did not corroborate, the essay is a primary account by the person named — while still leaving the central question of the entry untouched. He describes the culture and his reasons; he does not say what the sensitive information was, who received it, or who at OpenAI decided the three should go.