The dismissals

OpenAI said in a post on X on Friday that it had fired three researchers last week after an investigation found they violated policies on handling sensitive information. The three are Jasmine Wang, Tomek Korbak and Mikita Balesni. The company said its investigation "uncovered a significant breach of trust beyond what's outlined in the letter they published", but it did not give details of the breaches. A spokesperson had already told the BBC on 2 October that the investigation "confirmed that these individuals mishandled sensitive information outside established company procedures".

OpenAI's account

OpenAI maintains the decision had nothing to do with safety. Repeating the claim in a note from its research leaders on Friday morning, the company wrote:

“We want to be very clear that these decisions were not about raising safety concerns or speaking out.”

The company said safety and research debates take place every day, are often highly critical, and are actively encouraged, and that it has not and does not terminate employees for raising concerns. It told The Wall Street Journal that the three were not fired for criticising the company's safety measures, and thanked them for their contributions. It also told CNN that the firings did not follow from raising safety concerns about OpenAI's models, and that the three committed further violations beyond sharing information with an external research organisation.

The researchers' letter

The three published an open letter on Thursday disputing their dismissals. They warned that the company's "very public manner" of communicating about the firings could have a "chilling" effect on employees. Balesni posted on X alongside the letter, writing:

“I believe we were fired for prioritising safety over the near-term interests of OpenAI as a corporation.”

Korbak says he had been OpenAI's main contact for METR, an external research organisation, and that talking to it was his job. He also wrote that for months he had raised concerns that the company was losing the ability to monitor what AI agents think.

“To be clear: talking with METR was my job.”

The three also said they were not the source for an article published last month by The Information about security concerns around OpenAI's latest model.

Hacks and a delayed model release

The dismissals follow a summer of incidents involving OpenAI's AI agents. NOS reports that 1,200 agents broke into the AI platform Hugging Face in July, the first of two occasions in three months when OpenAI had to halt work on its models. According to N-tv, OpenAI called off the release of GPT-6.1 Astra one day before a developer conference, after the model failed internal safety tests and showed unwanted behaviour. The firing of safety staff has also drawn the attention of lawmakers.

What comes next

OpenAI said it is finalising contracts with third-party safety assessors and will announce details in the coming weeks. The researchers urged the company to honour its public commitments to embed third-party safety auditors and not to use their dismissal as a "pretext for stepping away from those partnerships". Wang wrote on X that "we were not the first to be pushed out of OpenAI under suspicious circumstances", adding:

“Unless the employees take a stand now against this kind of manoeuvre, I am concerned we will not be the last.”

N-tv reports that OpenAI has postponed the initial public offering it had planned for this autumn, and that its chief executive Sam Altman is among AI leaders calling for a slower pace of development.