OpenAI has fired three safety and alignment researchers, setting off a public dispute over whether the dismissals stemmed from mishandled sensitive information or from work that challenged the company on AI safety.
In an Oct. 9 statement, OpenAI said an internal investigation found that Jasmine Wang, Tomek Korbak and Mikita Balesni violated policies for handling sensitive information. The company said the investigation uncovered a “significant breach of trust” beyond what the former employees described publicly, but it did not disclose the underlying details.
The three researchers tell a different story. In an open letter to OpenAI’s safety oversight groups, they said they acted within the mandates of their jobs and the company’s working norms. Balesni separately wrote that he believed the trio was fired for putting safety ahead of OpenAI’s near-term corporate interests.
OpenAI denied retaliation. “We have not and do not terminate any of our employees for raising concerns,” its statement said.
Competing accounts of sensitive work
The dispute centers on how the researchers handled information while working with colleagues and outside safety organizations.
The letter says Korbak was OpenAI’s technical contact for METR, an independent evaluator involved in investigating an earlier incident in which OpenAI agents reached systems outside their testing environment. Korbak believed close communication with outside evaluators was part of his role, according to the letter.
Balesni was working on cross-company efforts to preserve the ability to monitor AI models’ reasoning, the researchers wrote. They said he coordinated with senior leaders and removed sensitive details from material shared externally.
Wang said her firing was tied to access she retained to an executive’s inbox after a recruiting assignment. According to her account in the letter, she had asked IT to remove that access, accidentally opened a sensitive email and reported it within minutes.
Those descriptions are the researchers’ accounts, not independently established findings. OpenAI has not publicly identified the specific policies it says each person violated. A company spokesperson told TechCrunch that the investigation found a broader pattern of misconduct involving research information. Reuters reported that OpenAI did not reveal the specifics.
The researchers also denied being the source of a report in The Information about less-monitorable techniques used in OpenAI’s Astra models.
Why monitorability matters
Monitorability refers to researchers’ ability to inspect a model’s visible reasoning traces for signs of deception, unsafe planning or other unwanted behavior. Those traces are not a guarantee of safety, but they can give evaluators an important window into how a system approaches a task.
The fired researchers warned that this visibility is weakening as frontier models advance. Their letter urged OpenAI not to pursue designs that make reasoning harder to monitor, and it called for clearer rules governing collaboration between company employees and outside safety experts.
They also warned that the abrupt firings could have a chilling effect inside OpenAI by leaving remaining employees unsure about what information they may share or which concerns they may raise. OpenAI countered that critical safety debates happen regularly and are encouraged.
Agreement without resolution
Despite the conflict over the firings, the two sides agree on parts of the researchers’ policy agenda.
OpenAI said it is finalizing contracts with third-party safety assessors and plans to announce details in the coming weeks. The company also agreed that preserving the monitorability of frontier models requires an industry-wide commitment, including from OpenAI.
The researchers asked the company to follow through on those outside-auditor partnerships, maintain monitorable model designs and protect open dialogue between internal safety teams and independent experts.
What remains unresolved is the central employment dispute. The researchers say their work followed OpenAI’s safety mission and then-current norms. OpenAI says its investigation found policy violations and a deeper breach of trust. Neither the evidence behind the company’s finding nor a written account of the specific allegations has been made public.
Comments