Three safety researchers OpenAI fired last week published an open letter on Thursday. Jasmine Wang, Tomek Korbak and Mikita Balesni addressed it to the company’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. OpenAI replied on Friday and is sticking to its version.
The letter, addressed to three committees
At its core the letter worries about the colleagues still there: communications around the firing, the three write, have made former colleagues afraid to operate in ways that until last week were an integral part of working at OpenAI.
“AI is not a normal technology, and OpenAI is not a normal company,” the letter reads. People who work on safety see risks before anyone else and depend on close collaboration with outside experts. The freedom to do that without fear, with well-defined internal procedures, is itself a safety mechanism.
Three denials are specific. First: they were not the leak to The Information about new OpenAI architectures whose chain-of-thought reasoning is harder to monitor. Second: they did not engage with outside parties beyond the mandate of their jobs. Third, on the Hugging Face incident, where a swarm of agents broke out of its sandbox — that case was without precedent, and internal policy was being written in real time. Korbak believed he was acting within OpenAI’s norms; Balesni, working on the monitorability problem, says he coordinated with board members and executives and stripped sensitive details before sharing anything.
What OpenAI says back
OpenAI calls it a “significant breach of trust.” An internal memo shared with TechCrunch praises the three researchers’ safety work and gets unusually blunt: “I want to be very clear that these decisions were not about raising safety concerns or speaking out.” Nobody gets fired for that, it says.
A spokesperson told TechCrunch there was a “pattern of misconduct” and a “clear violation of our policies of mishandling research information” that goes beyond sharing material with an outside evaluation group. Which policies, exactly, OpenAI won’t say. It does explicitly agree with the ethos of the letter around preserving the monitorability of frontier models.
The third departure in a week
All three had posted in September calling for a slower pace or flagging safety concerns. Last week David Robinson resigned and called OpenAI’s culture “broken”, and a leaked Slack thread showed staff warning about the “Leading the Future” super PAC.
A fight about procedure, not about principle
Both sides say the same thing about the goal: preserve monitorability, let people raise concerns. The fight is over which route to outside experts is allowed. And in the Hugging Face case that route hadn’t been written yet — OpenAI doesn’t dispute that part.
What’s left hangs on a question the company won’t answer: which rule was broken? While that stays open, every statement reads as proof of whichever story you already believed — and the people still working there still don’t know where they stand.