Tomek Korbak, Jasmine Wang and Mikita Balesni say OpenAI fired them over work they understood to be authorized. The company disputes that account, saying its investigation found a broader pattern of mishandling research information. It denies dismissing the researchers for raising safety concerns.
In an open letter described in TechCrunch’s October 8 reporting, the three warn that communications surrounding their firings have made former colleagues afraid to speak openly or collaborate with outside safety organizations.
At issue is whether OpenAI enforced established confidentiality rules or punished conduct that researchers reasonably believed fell within their responsibilities. The available reporting does not resolve that question: the underlying investigation records and relevant internal policies have not been published. The claimed chilling effect also remains the researchers’ assessment, not an independently established finding.
OpenAI Says Its Investigation Found More Than External Sharing
According to TechCrunch, OpenAI dismissed the three researchers the week before their October 8, 2026, public statements. The reported allegations concerned sensitive company information and its handling outside established procedures, including sharing with a third-party AI safety organization.
A spokesperson told TechCrunch that an investigation found a “pattern of misconduct” involving research information, in clear violation of company policies. OpenAI said the findings went beyond sharing information with an outside AI evaluation group.
A defense of external collaboration would not, by itself, answer every possible allegation about access to or handling of confidential material. The company’s assertion of a pattern of misconduct, however, does not establish what happened without the supporting findings.
OpenAI also shared an internal memo with TechCrunch, attributed to a research leader, that praised the researchers’ safety contributions and rejected retaliation as the explanation for their departures.
“I want to be very clear that these decisions were not about raising safety concerns or speaking out,” the memo reads. “We have always encouraged that and always will. We do not terminate employees for raising concerns.”
OpenAI’s response to TechCrunch did not specify which policies the researchers allegedly violated. Nor did it directly answer the publication’s questions about the circumstances of the dismissals or protections for employees who raise safety concerns and work with external evaluators. Without those details, readers cannot assess the company’s central claim against a defined rule and a documented sequence of events.
The Researchers Say Their Work Fit Their Mandates
The researchers addressed their letter to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. They argue that the dismissals reflect shifting expectations around work that depended on outside expertise.
They deny being the source of a leak to The Information concerning less-monitorable model architectures. They also say they did not believe their engagement with external parties exceeded their job mandates.
These denials address different questions. Rejecting responsibility for a particular leak does not settle a broader allegation about information handling. Their account of authorized collaboration directly challenges whether the boundary between permitted safety work and misconduct was clear at the time.
The letter describes Korbak’s communication with outside safety evaluators during an investigation involving AI agents and Hugging Face. In the researchers’ account, the incident was unprecedented and internal policies were being developed in real time. Korbak believed close communication with evaluators was consistent with OpenAI’s policies and norms and necessary to build trust.
If procedures were still taking shape, a fair assessment would need to distinguish an established restriction from an expectation clarified later. The public reporting does not provide enough information to determine which applied.
Balesni’s account concerns work on model monitorability: the ability to examine a model’s reasoning for signs of problems. The letter says his effort required extensive communication with external parties and was coordinated with, and supported by, OpenAI board members and executives. The researchers say he checked with his reporting line and removed sensitive details before sharing materials. They describe his conduct as good-faith work within the company’s norms at the time.
Those claims bear on whether he had authorization, though the available evidence does not independently verify them. Approval to collaborate can be narrower than approval to share a particular document or detail. Establishing what was authorized would require the communications, permissions and materials at issue; a general statement that external work was encouraged would not settle the question.
Wang Says Delegated Email Access Was Not Removed
Wang offers a separate explanation for her dismissal. In her public statement, she says OpenAI gave her one reason: accessing an executive’s email.
In the fuller account quoted by TechCrunch, Wang says OpenAI delegated that access to her for recruiting. When she no longer needed it, she asked IT to remove it. She says the request was not acted on, she could not remove the access herself, and the executive’s inbox appeared combined with her own in her phone’s mail application.
Wang says she opened a sensitive email accidentally, informed the executive within minutes and again asked IT to remove the access. She argues that the access originated in an assigned task and that she disclosed the incident instead of concealing it.
Having access and being authorized to read a particular message are different things. Delegated access could explain why an email was available without resolving whether opening it complied with policy. Her account of an accidental opening and prompt disclosure is material, but the public evidence does not establish the relevant permissions or access history.
There is also an unresolved mismatch between the accounts. Wang describes being given one specific reason, while OpenAI publicly refers to a pattern extending beyond external information sharing. The reporting does not establish whether that broader description applies identically to each researcher or encompasses different allegations against different people.
Wang has called for a written, complete list of allegations so the researchers can address them. Such a list would not necessarily be made public, but it would clarify whether the two sides are disputing the same conduct.
Clear Permissions Are Part of Effective Safety Work
The researchers’ most consequential claim concerns employees who remain at OpenAI. They say conduct previously understood as normal now appears capable of triggering termination, leaving colleagues unclear about where they stand.
Their letter describes freedom to collaborate with outside experts, supported by well-defined internal procedures, as an “essential safety mechanism.” Researchers may need external scrutiny to understand risks, while the company still needs controls over confidential information.
