Tomek Korbak

The researchers say the firings are scaring staff away from raising safety concerns, while OpenAI says an investigation found they broke clear rules on handling sensitive information.

 

OpenAI fired three of its AI safety researchers, and the two sides now give opposing accounts of why. The three, Tomek Korbak, Jasmine Wang, and Mikita Balesni, published a letter on Oct. 8 saying their firing has made colleagues afraid to raise safety concerns or to work with outside safety groups. Balesni wrote on X that he believes they were fired “for prioritizing safety over the near-term interests of OpenAI as a corporation.”

OpenAI answered on Oct. 9. Its research leaders said an internal investigation found the three “violated clear policies on handling sensitive information,” and that the firings “were not about raising safety concerns or speaking out.”

Either account describes a problem inside one of the companies building the most powerful AI models.

The three researchers worked on spotting dangerous AI behavior

The three say they each worked on AI safety for years. At OpenAI, much of their work involved checking how AI models behave, including by reading the step-by-step reasoning they write as they work.

Korbak and Balesni both worked on OpenAI’s investigation into the Hugging Face break-in. During the investigation, Korbak was OpenAI’s technical contact for METR, a nonprofit that tests AI systems for dangerous behavior.

The researchers say they stayed within OpenAI’s rules

The letter is addressed to three OpenAI groups that oversee its safety work. In it, the three say they acted within the company’s rules as they stood at the time, and that conduct considered normal last month now appears to be grounds for firing.

The letter responds to different versions of events regarding the firings. OpenAI has not said what the three did beyond breaking its rules on sensitive information, so the letter is the only public account of the conduct in question.

A leak. The three say they were not the source of a story in The Information about new OpenAI model designs whose reasoning would be harder to read. They say that story undercut their work with other AI companies on shared limits for building such designs.

Work with outside groups. Korbak and Balesni say their contact with outside groups was part of their jobs. Balesni says he discussed his work with other AI companies with OpenAI board members and senior executives.

An executive’s email. Wang says she was given access to an executive’s email to help with recruiting. She asked IT to remove that access once she no longer needed it, and IT did not. When she accidentally opened a sensitive email, she says she told the executive within minutes.

A board memo. The letter mentions rumors about a board-level memo. The three say no one raised it with them, so they had no chance to respond.

OpenAI says its investigation found more than the letter describes

OpenAI’s reply, posted on its newsroom account on X, is signed only as “a note from our research leaders.” It says OpenAI’s internal investigation “uncovered a significant breach of trust beyond what’s outlined in the letter they published.” It does not say what that breach of trust was.

The reply says OpenAI generally keeps employment matters private, and that the company has not fired, and does not fire, employees for raising concerns.

The researchers want outside safety reviewers kept inside OpenAI, and OpenAI says contracts are close

The letter ends with three requests. First, OpenAI should keep its promise to bring in outside safety reviewers. They work for independent groups such as METR, under contract, but get access to OpenAI similar to an employee’s. According to the letter, Sam Altman made that promise publicly on Sept. 12. The three worry OpenAI could point to the firings as a reason to end its work with METR, or to give outside reviewers less access. Anthropic CEO Dario Amodei made a similar pledge in September.

Second, OpenAI should not build models that make it harder for staff to check their reasoning. Today, staff can read the step-by-step reasoning a model writes out as it works, and the letter says companies rely on that to check for safety problems. The letter says this reasoning is getting harder to read across the industry, and that no company yet knows how to safely release a model whose reasoning it cannot check. It does not say why.

Third, OpenAI should set clear rules for how employees may work with outside safety groups, “so that no one has to guess where the shifting lines now are.”

OpenAI’s reply agrees that keeping model reasoning readable needs a commitment from the whole industry, including OpenAI. It says the company is finalizing contracts with outside safety assessors and will announce details in the coming weeks.