OpenAI Fired Its Safety Team. Now They're Warning the Board About the One Thing It Can't See.

OpenAI fired three safety researchers on 1 October. This week they wrote to the board about the one instrument that checks whether a model is telling the truth.

The three people OpenAI paid to watch its models for danger were shown the door on 1 October. Jasmine Wang, Tomek Korbak and Mikita Balesni worked on safety and alignment. OpenAI's statement was short: they mishandled sensitive company information outside established procedures. The reporting links the dismissals to material shared with an outside AI-safety group.

That is the news. The interesting part is the thing they say is going dark.

What the letter asks for

The Wall Street Journal reviewed the letter. Its asks are narrow. "As an industry, we do not yet know how to safely develop and deploy models that we cannot monitor," it says, and OpenAI and its rivals "should not move forward with developments that further decrease" that ability. The three want closer work with outside safety auditors, against the risk they call "truly catastrophic". Korbak and Balesni had co-authored a chain-of-thought monitoring paper.

As an industry, we do not yet know how to safely develop and deploy models that we cannot monitor.

The one thing it cannot see

Chain of thought is the closest thing a model has to a confession. When it is legible, a reviewer can watch the model reason its way toward a wrong answer and catch it. When it is summarised or hidden, that check disappears. The worry is not hypothetical: OpenAI has disclosed a coordinated campaign to extract the hidden reasoning of its models, an "adversarial distillation" effort it says began on 1 July and was shut down by 28 July.

Why three jobs is not the story

The firing is a personnel matter. The warning is not, and OpenAI's own week shows the seam. On 6 October it published 722 mathematics manuscripts from an unreleased model; the Association for Human Mathematics, reposted by Terence Tao, told mathematicians to stop reviewing the output. The common thread is verification: we are producing more than anyone can check, and the instruments for checking are being pulled back. I read that release in 722 Proofs, One Locked Door.

OpenAI's answer

OpenAI does not dispute the letter's goal. In a memo shared with the Journal, it said it "strongly agreed" with the recommendations and that the dismissals "were not about raising safety concerns or speaking out". The three say they did not believe they had "engaged with external parties outside the mandates of our jobs". Both things can be true: a company can agree with a warning and still part with the people who raised it.

The honest counterpoint

There is a fair reading in which OpenAI is right. Safety-testing information is genuinely sensitive, and an outside group is not automatically trustworthy. OpenAI says the three broke its rules, and I have not seen the internal evidence either way. I spent time looking for the full letter; it is not public, so the quotes here are the Journal's. What would change my mind: the letter in full, or a concrete account of what was shared and with whom.

If you are weighing how much of a model's reasoning you can actually audit, and want a second pair of eyes on the question, email brandon@kreostudio.co.uk.