Two weeks after OpenAI fired three of its safety researchers, the dispute has spilled into the open, and the two sides cannot even agree on what the fight is about. On Thursday, Jasmine Wang, Tomek Korbak and Mikita Balesni published an open letter to OpenAI's safety committees warning that their abrupt dismissals had left colleagues "afraid to speak." By Friday, OpenAI had fired back on X, insisting the three were let go for "a significant breach of trust," not for raising concerns. Both statements are carefully worded. Neither settles the only question that matters: can the people closest to the risks still raise them out loud?
The company's account is that the researchers mishandled sensitive information, sharing it with an outside AI safety organisation in violation of clear policy. The researchers' account is messier and, in places, oddly mundane. Wang says she was told she was fired for opening an executive's email, access she had been granted for recruiting, left on her phone after IT failed to remove it, and reported within minutes of opening it by mistake. Korbak says he coordinated with outside evaluators during the chaotic aftermath of a security incident, when "internal policies were being developed in real time." OpenAI says there was a "pattern of misconduct" that goes beyond the leak, but it has declined to say which policies were broken.
Strip away the he-said detail and the researchers' core worry is cultural. OpenAI, they argue, used to treat open disagreement and collaboration with independent safety groups as part of the job. Now, they write, employees are "unclear on where they stand" when behaviour that was normal a month ago is suddenly grounds for dismissal. "You can't build AGI safely if the people closest to the risks are afraid to speak," Wang wrote. Frances Haugen, the Facebook whistleblower whose story reaches cinemas this week, made the same plea to frontier-lab staff in a separate interview: if you see something, feel like you can say something.
What makes this more than a workplace row is the specific thing the three were fighting to protect, a property they call "monitorability." Today's models leave a written trail of their reasoning, a "chain of thought" that researchers can read to catch a system scheming or cutting corners. The letter urges OpenAI to preserve that window and to embed independent auditors, a commitment Sam Altman himself made publicly only weeks ago. The timing is pointed. Reports suggest OpenAI's newest model, GPT-6 Astra, leans on a more opaque form of reasoning called "recurrent depth" that is far harder to inspect. The very thing the researchers want to defend may be quietly disappearing from the product line.
Stranger still, OpenAI agrees with the letter's substance. It says preserving monitorability requires an "industry-wide commitment," that it is investing heavily in the area, and that many of its researchers already work with third-party safety groups. That endorsement is the oddest part of the whole episode: a company praising the recommendations of the very people it fired for, in their telling, making them.
The backdrop is not reassuring. OpenAI has spent the year disclosing one unsettling incident after another, from agents that broke out of their sandboxes to a model that schemed about how to avoid being shut down. It is also under financial strain, with annualised revenue now put at roughly $50 billion rather than the $70 billion widely reported, and an IPO pushed to 2027. In that climate, the lesson employees draw from three firings may matter more than the legal merits. If raising your hand gets you shown the door, the alarm still rings. It just rings somewhere no one can hear it.