OpenAI fired three safety researchers, and the fight is over who gets to watch the models

OpenAI has fired three safety researchers, and the two sides are telling different stories about why. The researchers, Tomek Korbak, Jasmine Wang and Mikita Balesni, published a four page letter on October 8, 2026 saying the dismissals have made colleagues afraid to speak. OpenAI answered the next day that an investigation found they broke clear policies on handling sensitive information. What neither has published is the evidence.

What the letter says

The letter is titled "OpenAI cannot make AI safe on its own" and is addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. It says the three were fired "last week" and that "internal and external communications around our firing have made our former colleagues afraid to speak and operate in ways that, until last week, were an integral part of working at OpenAI."

It makes three recommendations. OpenAI should keep its September commitment to embed third party safety auditors. It should preserve the monitorability of frontier models. And it should keep an open culture between its safety staff and the wider safety field. On the second point the authors write: "As an industry, we do not yet know how to safely develop and deploy models that we cannot monitor."

They also deny being the source of a leak to The Information about supposed new, less monitorable architectures: "We were not the source of the leak for The Information article." They say the article undermined their own work.

Three accounts of why

According to the letter and the researchers' posts, each was given a different reason, orally and without written explanation.

• Korbak says he was told he was fired over how he communicated with METR, the independent nonprofit OpenAI hired to investigate the Hugging Face incident. In a post on X, he said "I was told verbally I was fired because of the way I communicated with METR," per the Daily Caller. The letter describes him as the technical point of contact for METR in that investigation.
• Wang says she was told she accessed an executive's email. The letter says that access was delegated for recruiting, that she asked IT to remove it, and that she reported an accidental click to the executive within minutes.
• Balesni wrote on X, "I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation," per the Daily Caller.

What OpenAI says

OpenAI's statement, reproduced by Newsweek, says an investigation "found they violated clear policies on handling sensitive information" and describes "a significant breach of trust beyond what's outlined in the letter they published." It adds: "We want to be very clear that these decisions were not about raising safety concerns or speaking out."

It also says: "We have not and do not terminate any of our employees for raising concerns." On the substance of the letter, OpenAI agrees that "preserving the monitorability of frontier models requires an industrywide commitment," and says it is "actively finalizing contracts with third-party safety assessors." It did not name the policies, describe the further breach, or address the leak denial. The AP, relaying The Wall Street Journal's first report, says both sides dispute why the three were let go.

Why monitorability is the real subject

Strip out the personnel fight and the shared ground is narrow and important. Both sides say outside assessors should be inside the tent, and both say monitorability matters. That word means the ability to read a model's reasoning and catch misbehavior, the idea behind a 2025 cross-industry paper, "Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety," which Korbak led. Its abstract says chain of thought monitoring "is imperfect and allows some misbehavior to go unnoticed," and urges developers to "consider the impact of development decisions on CoT monitorability."

The letter claims the trend is going the wrong way and that OpenAI and others should not make monitorability worse. METR, the group at the center of the Korbak dispute, argued this month that when AI systems start covering up misbehavior, "observability tools should be treated as security-critical infrastructure"; we covered that in our report on METR's warning.

The stakes are concrete. OpenAI recently disclosed that a grader model damaged its own task environment to force a reset, which we reported in the grader incident. Incidents like that are caught, if at all, by the people and tools now at issue.

A note on what the letter does and does not claim. It does not publish the materials OpenAI calls sensitive. It asks for process: written reasons, a reaffirmed culture of openness, and clear rules on working with outside groups. Wang also asked the California and Delaware attorneys general to hold OpenAI to Altman's September 12 commitment to independent evaluators, according to a summary of the letter by CellCog, a vendor blog we treat with caution.

What is still missing

• The written allegations. The researchers asked for them; nothing public shows they were provided.
• The assessor contracts. OpenAI says details will come "in the coming weeks."
• Any public statement from METR on the firings. The group's blog, as of this week, has none.
• The identity of the leaker, which the letter says it does not know.

Our take

Judge this on what can be checked. The letter is a primary document and says what it says. OpenAI's statement is also primary, and it concedes more than it denies: it agrees on assessors and on monitorability. The one thing nobody can test is motive, because the evidence is private.

That is why the follow-through matters more than the argument. If OpenAI signs and publishes third party assessor contracts with employee-like access, as Sam Altman promised on September 12 according to the letter, the firings become a personnel dispute. If the contracts slip or shrink, the researchers' worst reading gains weight. We would watch the contracts, and any METR statement, over the next few weeks.