Mikita Balesni, Jasmine Wang and Tomek Korbak, who worked on safety and alignment research at OpenAI, published an open letter on October 8, 2026, U.S. time, saying they had been dismissed the previous week. They said they feared the company’s ability to discuss safety concerns openly and work with external experts was being weakened. OpenAI said the dismissals were due to violations of its rules for handling confidential information.

The letter was addressed to three bodies overseeing OpenAI safety: the Safety and Security Committee, the Safety Advisory Group and the Mission Advisory Council. The researchers said their communications with outside organizations were part of their work. Korbak wrote on X that he was verbally told his communication with external evaluator METR was the reason for his dismissal, but was not given the specific statements or actions at issue, when they occurred, or a written explanation.

Wang said access to executives’ emails, which she had been authorized to use for recruiting, remained in place after she asked for it to be revoked. She said she reported to an executive within minutes after accidentally opening a confidential email. The three also denied being the source of a leak for a report by The Information about model designs that make AI reasoning harder to monitor.

Reuters reported on October 9 that OpenAI said an internal investigation had found the three violated its confidential-information handling rules. The company said there had also been a serious breach of trust beyond the conduct described in the letter, and that it does not dismiss employees for raising safety concerns. OpenAI has not disclosed the specific violations.

The researchers called for continued cooperation with external safety auditors, preservation of the monitorability of frontier models, and protection of a culture that allows candid discussion of safety. Monitorability is the ability to detect signs of improper behavior from an AI system’s reasoning process, among other signals. A research paper co-authored by the three said monitoring reasoning processes is a promising safety measure, but remains imperfect and can be undermined by development choices.

OpenAI also reported in the official system card for GPT-6 Astra that its reasoning process was less monitorable than that of earlier models. The company said it checks model safety through multiple evaluations and safeguards, but no causal link has been established between the reduced monitorability and these dismissals. The dispute centers on the researchers’ account, OpenAI’s explanation and the actions requested in their letter.

Mikita Balesni, Jasmine Wang and Tomek Korbak, who worked on safety and alignment research at OpenAI, published an open letter on October 8, 2026, U.S. time, saying they had been dismissed the previous week. They said they feared the company’s ability to discuss safety concerns openly and work with external experts was being weakened. OpenAI said the dismissals were due to violations of its rules for handling confidential information.

The letter was addressed to three bodies overseeing OpenAI safety: the Safety and Security Committee, the Safety Advisory Group and the Mission Advisory Council. The researchers said their communications with outside organizations were part of their work. Korbak wrote on X that he was verbally told his communication with external evaluator METR was the reason for his dismissal, but was not given the specific statements or actions at issue, when they occurred, or a written explanation.

Wang said access to executives’ emails, which she had been authorized to use for recruiting, remained in place after she asked for it to be revoked. She said she reported to an executive within minutes after accidentally opening a confidential email. The three also denied being the source of a leak for a report by The Information about model designs that make AI reasoning harder to monitor.

Reuters reported on October 9 that OpenAI said an internal investigation had found the three violated its confidential-information handling rules. The company said there had also been a serious breach of trust beyond the conduct described in the letter, and that it does not dismiss employees for raising safety concerns. OpenAI has not disclosed the specific violations.

The researchers called for continued cooperation with external safety auditors, preservation of the monitorability of frontier models, and protection of a culture that allows candid discussion of safety. Monitorability is the ability to detect signs of improper behavior from an AI system’s reasoning process, among other signals. A research paper co-authored by the three said monitoring reasoning processes is a promising safety measure, but remains imperfect and can be undermined by development choices.

OpenAI also reported in the official system card for GPT-6 Astra that its reasoning process was less monitorable than that of earlier models. The company said it checks model safety through multiple evaluations and safeguards, but no causal link has been established between the reduced monitorability and these dismissals. The dispute centers on the researchers’ account, OpenAI’s explanation and the actions requested in their letter.