OpenAI has fired three researchers after an internal investigation found they violated the company’s policies on handling sensitive information, the company said in a statement reported this week. The dismissed researchers — Jasmine Wang, Tomek Korbak and Mikita Balesni — had published a letter describing their terminations and expressing concern about how the matter was handled.
In its statement, OpenAI said the investigation “uncovered a significant breach of trust” beyond what the researchers’ letter described, and that it stood by the decision to end their employment. The company was at pains to draw one line clearly: it asserted the firings were not punishment for raising safety concerns or speaking critically, saying such debate happens daily inside OpenAI and that no employee is terminated for it.
The researchers, for their part, said in their letter that they were not the source of a news article published last month about security concerns surrounding OpenAI’s latest model — the leak that appears to have prompted the investigation. Their letter warned that abrupt dismissals of this kind could chill the internal culture that allows safety concerns to surface, an argument that lands with particular force at a company whose founding mythology is built on the importance of raising exactly those concerns.
Both accounts cannot be fully evaluated from outside: the company will not detail the evidence, and the researchers cannot publish it. What can be said is that the episode illustrates the tightening information regime inside frontier AI labs, where model details carry enormous commercial and strategic value and where the line between whistleblowing, leaking and legitimate internal dissent is being negotiated case by case — publicly, and with reputations attached.
The departures arrive as OpenAI prepares for a widely expected public listing in early 2027, when its internal controls will face a far less forgiving audience than the technology press.
The episode will be studied as a marker of the industry’s maturation — the point at which frontier labs adopted the information disciplines of defence contractors while still describing themselves in the language of open research. That tension is genuinely unresolvable by memo: the same openness that accelerates safety science also arms competitors, and the same secrecy that protects a model also starves the scrutiny that might catch its dangers. Every major lab is improvising its answer. OpenAI’s, this week, was delivered by termination letter, and the debate it intended to close has instead been given three names, a published letter and a permanent place in the company’s history.
