Three fired OpenAI safety researchers are challenging the company’s explanation for their dismissals, warning that uncertainty over what employees can share with outside experts could undermine safety work.
Jasmine Wang, Tomek Korbak, and Mikita Balesni deny mishandling sensitive information outside the scope of their jobs. In an open letter addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council, they argue that their dismissals could make former colleagues reluctant to raise concerns or pursue external collaboration.
OpenAI rejects the suggestion that the researchers were fired for speaking up about safety. The dispute leaves an unresolved question at the center of their competing accounts: whether the researchers crossed established boundaries or were punished for work they understood the company to permit.
OpenAI and the researchers offer conflicting explanations
OpenAI alleges that the three mishandled sensitive company information. A company spokesperson characterized the findings of an internal investigation as a pattern of misconduct extending beyond sharing information with an outside AI evaluation group. Those allegations have not been independently verified, and the researchers dispute the company’s characterization of their conduct.
The researchers also deny that they engaged with external parties beyond the mandates of their jobs. That distinction matters to their defense: they describe contact with outside safety specialists as part of the work OpenAI employed them to do.
An internal memo attributed to an OpenAI research leader praised their contributions to AI safety while rejecting the claim that the dismissals were retaliatory. The memo presented raising concerns and speaking openly as activities the company encourages.
That statement establishes OpenAI’s position, but it does not settle the disagreement over how its policies applied to these employees. OpenAI did not provide specific answers about which policies the researchers allegedly violated, the circumstances of their dismissals, or the protections available to employees collaborating with external evaluators.
Without those details, the broader misconduct allegation and the researchers’ account of authorized work remain difficult to reconcile. Neither account, by itself, establishes the full circumstances behind the decisions.
The warning about a chill on safety work
The researchers’ central concern extends beyond their own employment. They argue that the way the dismissals were handled and communicated could change how people inside OpenAI approach uncomfortable questions.
Their letter describes a workplace that previously encouraged employees to raise safety concerns and disagree openly. In their view, colleagues have been left uncertain about whether conduct they understood to be acceptable could expose them to dismissal. That is their assessment of the company’s culture, rather than an established finding about how all employees feel.
For Wang, Korbak, and Balesni, outside collaboration is part of the safety process itself. They argue that researchers who encounter emerging risks need access to other experts to understand those risks and decide how to address them. Clear internal procedures, in their account, help make that collaboration possible.
The concern is therefore about both permission and predictability. Their argument is that employees need to understand where the boundaries are before a sensitive exchange takes place. A general commitment to open discussion offers limited reassurance if researchers cannot tell how the company will interpret a particular interaction with an outside evaluator.
External collaboration sits at the center of their defense
The letter presents different explanations for the work Korbak and Balesni were doing. Both accounts emphasize communication with people outside OpenAI, but they concern separate safety efforts.
Korbak’s defense concerns his contact with external safety evaluators during a sensitive investigation. He believed that close communication helped build trust and was consistent with OpenAI’s policies and norms. The researchers describe the investigation as unprecedented and say internal procedures were being developed as it unfolded. Those descriptions of the investigation and its evolving rules have not been independently verified.
The distinction they draw is between deliberately disregarding a known restriction and acting within what an employee understood the rules to allow. Their letter argues that Korbak was doing the latter; OpenAI’s broader misconduct allegations do not publicly resolve that specific disagreement.
Balesni’s work concerned model monitorability, including the difficulty of examining chain-of-thought reasoning. The researchers argue that addressing this problem requires extensive communication with external parties.
They say Balesni coordinated with OpenAI board members and executives, checked with his reporting line, and removed sensitive details before sharing materials. These are elements of the researchers’ defense, rather than independently established conclusions about whether every exchange complied with company policy.
The three also deny involvement in a leak about model architectures that could make reasoning harder to monitor. Their denial separates that allegation from the external collaboration they acknowledge and defend as part of their responsibilities.
Wang disputes the explanation involving an executive’s email
Wang offered a separate account of her dismissal in a post on X. She said OpenAI told her she had been fired because she accessed an executive’s email, and she disputed the implication that the access amounted to misconduct.
According to Wang, OpenAI had delegated access to that mailbox to her for recruiting. She said she asked IT to remove it when she no longer needed it, but the request was not carried out and she could not revoke the access herself.
Wang also described a phone mail app that combined the inbox with her other messages in a way she could not distinguish. She said she opened a sensitive email accidentally, informed the executive within minutes, and again asked IT to remove her access. Her account of the mailbox configuration, removal requests, and notification sequence has not been independently verified.
Her defense turns on the circumstances of the access: why it was granted, why it remained available, and what she did after opening the message. She maintains that her actions were not hidden and questions whether the reasons given for the three dismissals adequately explain what happened.
The researchers want clearer rules and outside accountability
The letter calls on OpenAI to uphold its public commitments to embedding third-party safety auditors within the organization and preserving the monitorability of frontier models. It also asks the company to support open dialogue between its safety researchers and the wider safety community.
The internal memo expressed agreement with those recommendations. That suggests some alignment on the stated goals, even as the two sides remain divided over the dismissals and the conduct that preceded them.
For the researchers, the practical issue is whether those commitments give employees enough clarity to act. They want external collaboration to remain a recognized part of safety work, with procedures that allow researchers to understand what information they can share and under what conditions.
Wang’s broader warning is that fear of dismissal could discourage the people closest to potential risks from speaking openly. That is a concern about the possible consequences of the firings, not proof that every employee has been silenced. Resolving the dispute would require addressing both the specific misconduct allegations and the uncertainty the researchers say their departures have created.