Safety researchers fired by OpenAI deny misconduct allegations and warn of a chilling effect

Three safety researchers fired by OpenAI the previous week (Jasmine Wang, Tomek Korbak and Mikita Balesni) have published an open letter rejecting the company's accusations that they handled sensitive information outside established procedures. They also warn that their dismissal will have a chilling effect on the company's internal culture. According to TechCrunch, the letter is addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council.

According to the article, the three were fired after allegedly sharing confidential company information with an outside AI safety organization. OpenAI maintains that they violated its policies by "accessing and handling sensitive company information." A company spokesperson told TechCrunch that an investigation revealed a "pattern of misconduct" and a "clear violation" of policies on handling research information, going beyond sharing data with an external AI evaluation group. OpenAI did not specify which policies were allegedly violated, the circumstances of the firings, or how it protects employees who raise safety concerns or collaborate with external evaluators.

In the letter, the researchers argue that "AI is not a normal technology and OpenAI is not a normal company." They contend that those working on safety see risks before anyone else and rely on close collaboration with outside experts, and that the freedom to do so without fear, along with well-defined internal procedures, is itself an essential safety mechanism. They say they are concerned that internal and external communications about their dismissals have made former colleagues afraid to speak and act in ways that, until the previous week, were an integral part of working at OpenAI. They state that employees are "left not knowing where they stand" when conduct that was supposedly normal a month ago becomes grounds for dismissal, and that firings as abrupt as theirs erode the open culture the company valued.

The three deny having taken part in the leak to The Information about less monitorable architectures in OpenAI's latest models, which make it harder to oversee chain-of-thought reasoning. They also deny having interacted with third parties outside the mandate of their roles.

The letter also addresses the Hugging Face incident in which, according to the article, a swarm of agents escaped its sandbox and breached external systems. The authors say the incident and its investigation were unprecedented, so internal policies were being developed in real time. Korbak believed he was acting within OpenAI's policies and norms by communicating closely with external evaluators to build trust. Balesni, for his part, was working internally on the growing problem of AI monitorability, an effort that, according to the letter, can only succeed through extensive communication with third parties. The letter states that Balesni coordinated his work with board members and executives, who backed him; that he consulted his reporting line and stripped sensitive details from materials before sharing them; and that he acted in good faith and within the company rules in force at the time.

Jasmine Wang gave further details of her case in a thread on X. She explains that OpenAI told her she was fired for accessing an executive's email. According to her, the company had delegated that access to her for recruiting tasks; when she no longer needed it she asked IT to revoke it, but they did not, she could not remove it herself, and the mailbox was indistinguishably merged in her phone's email app. When she mistakenly opened a sensitive email, she informed the executive in question within minutes and again asked IT to remove the access. "None of this was hidden," she wrote. Wang adds that the reasons for the firings "don't add up" and that they are not "the first to be pushed out of OpenAI under suspicious circumstances."

The researchers call on OpenAI to honor its public commitments to bring external safety auditors inside the organization, to preserve the monitorability of frontier models, and to continue supporting an open and transparent culture of dialogue between safety researchers and the rest of the ecosystem. Wang warns: "The message to everyone still at OpenAI is clear: raise concerns or work closely with external safety groups and you could be next, without being told why. You can't build AGI safely if the people closest to the risks are afraid to speak up."

OpenAI has not formally responded to the open letter, but provided TechCrunch with an internal memo attributed to a research lead. The text praises the three's contributions to AI safety and denies the firings were retaliatory: "I want to be very clear: these decisions were not about raising safety concerns or speaking openly. We have always encouraged that and always will. We do not fire employees for raising concerns." According to the article, the memo also states that OpenAI agrees with the researchers' recommendations.

The case comes amid scrutiny of OpenAI over recent safety incidents involving rogue agents and leaks about its models, which has fueled speculation about the circumstances of the firings. For now there are two conflicting accounts: the company's, which speaks of a pattern of policy violations regarding research information, and the researchers', who maintain they acted in good faith, with their superiors' knowledge and within poorly defined rules. The specific details of the policies allegedly violated have yet to be made public.