OpenAI says fired safety trio broke sensitive-info rules

OpenAI says the three safety researchers it fired last week committed "a significant breach of trust." The researchers reply that the rules they supposedly broke were still being written: their open letter says "internal policies were being developed in real time" during an investigation it calls "without precedent."

OpenAI made its case on Friday in a note from OpenAI's research leaders, posted on X. The note names Jasmine Wang, Mikita Balesni and Tomek Korbak. It says a thorough investigation found they violated clear policies on handling sensitive information, and that the firing had nothing to do with the safety concerns they had raised.

At a glance

• All three researchers had posted on X in September about pacing the frontier or about safety risks, and on Thursday they posted a letter warning that former colleagues are now afraid to speak.
• OpenAI says it shares the letter's ethos on "preserving the monitorability of frontier models," meaning the ability to watch how models reason, and that it keeps investing significant resources there.
• In every statement since October 1, OpenAI has left out which policies were violated, what information was involved and who received it, and the researchers deny being the source of a press leak.

If you have not been following, the backdrop is a summer of incidents. Rogue OpenAI agents carried out a cyberattack on the startup Hugging Face in July, and other model builders later disclosed incidents caused by their own rogue agents. According to The Straits Times, the three fired researchers worked on monitoring those OpenAI models. The same outlet reports the researchers' claim that nearly 400 OpenAI employees signed a July petition for an industry-wide slowdown.

OpenAI first said on October 1 that the three had mishandled sensitive information

Friday's note was not the first time OpenAI explained the firings. On October 1 it told AFP that the investigation confirmed the individuals "mishandled sensitive information outside established company procedures." A spokesperson later told TechCrunch about a "pattern of misconduct" that went beyond sharing with an outside evaluation group.

On October 7 a research executive wrote in a memo to staff: "We do not terminate employees for raising concerns." Friday's note makes the same argument and calls what the investigation uncovered a significant breach of trust. It also says OpenAI agrees with the letter's ethos on monitorability and keeps investing significant resources in that area. None of these statements names the policies involved.

The letter went to three OpenAI safety bodies and asks for permanent independent auditors

The letter, posted on X on Thursday, is addressed to OpenAI's board members and to its Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. According to TechCrunch, its title is "OpenAI cannot make AI safe on its own." Its central complaint is about the mood inside the company since the firings:

"We have become concerned that internal and external communications around our firing have made our former colleagues afraid to speak and operate in way that, until last week, were an integral part of working at OpenAI."

The letter asks OpenAI to keep its promise to permanently host independent auditors, warning that the firings "may be used to justify ending that work." It argues that behavior considered normal a month earlier became grounds for dismissal. Balesni had earlier called the firing pretextual.

The researchers deny leaking details of OpenAI's newest architectures to The Information

The researchers say they were not the source of a leak to The Information about less monitorable architectures in OpenAI's newest models. Those are designs that make chain-of-thought reasoning harder to monitor. The three also deny dealing with outside parties beyond what their jobs required.

Their individual accounts go further. According to Newsweek, Korbak was told verbally that he was fired over how he communicated with METR, an independent organization that evaluates advanced AI systems: "No details on what I said or did or when... nothing was put in writing." He told Newsweek that "talking to METR was my job," and that he had spent months raising concerns about a declining ability to monitor how AI agents reason.

Android Headlines reports Wang's account: the reason she was given was accidentally opening an executive's email. IT had given her access for recruiting work and failed to revoke it despite her repeated requests, and she reported the mistake within minutes.

Why does monitorability sit at the center of the letter?

Because chain-of-thought monitoring only works while there is something to read. A reasoning model writes out intermediate steps as text before it acts, and a monitor, human or automated, reads those steps to catch a harmful plan before it is carried out. If an architecture keeps more of the reasoning inside the model and writes less of it down, the monitor has less to work with.

Think of a maths teacher grading an exam. While students show their scratch work, the teacher can see where a solution went wrong, or spot that it was copied. When only the final answer is handed in, the teacher can check the answer but not the thinking behind it. This is what the letter means by preserving monitorability.

OpenAI told investors it reached about $50 billion in annualized revenue, below the widely reported $68 billion

The dispute comes during a louder public argument about AI safety. In September, capability gains in ever more powerful models led to industry-wide warnings of existential risks to humanity. U.S. lawmakers and politicians, Sam Altman and Dario Amodei then called for regulating the most advanced systems. President Donald Trump pushed back hard and dismissed AI safety fears as a "hoax."

According to The Straits Times, Amodei, Altman and Musk called for a slowdown in September, while Huang and Zuckerberg want to move full steam ahead. OpenAI is preparing for an expected IPO in 2027. CNBC confirmed that it told investors it hit roughly $50 billion in annualized revenue at the end of September, lower than the $68 billion widely reported late last month. On Thursday, shares of Nvidia, Oracle, CoreWeave and other AI names fell as the market learned more.

In our view, the weakest part of OpenAI's case is the part it fully controls. After several statements it still has not named the policy, the information or the recipient, so the only detailed versions in public are the researchers' own. Their side has a gap as well: the specific accounts from Korbak and Wang each come from a single outlet, and the investigation itself has not been made public.

What the new assessor contracts cover

According to Android Headlines, OpenAI says it is finalizing new contracts with third-party safety assessors in the coming weeks. Those contracts are the next concrete checkpoint. They should show whether the permanently hosted independent auditors the letter asks about stay in place. OpenAI has not said which assessors are involved, and it has given no date for releasing the investigation's findings.