Fired OpenAI researchers contest misconduct allegations and ask for clearer safety rules

Three safety researchers dismissed by OpenAI have challenged the company's misconduct allegations and warned that uncertainty over acceptable outside collaboration could discourage safety work. OpenAI disputes the retaliation claim and says an investigation identified a pattern of mishandling research information.

In Rebecca Bellan's reporting for TechCrunch, Jasmine Wang, Tomek Korbak and Mikita Balesni describe their concerns in an open letter to three OpenAI oversight groups. The account presents competing explanations for their dismissals. It does not resolve which interpretation of the company's rules is correct.

The researchers say collaboration rules became unclear

Wang, Korbak and Balesni say colleagues have become afraid to speak or work in ways that were accepted before the firings. Their argument links external collaboration to safety: researchers may need outside expertise to understand risks before the company releases a system. They call for well-defined procedures that allow that work without fear.

The letter denies involvement in a leak to The Information concerning model architectures whose chain-of-thought reasoning is harder to monitor. The researchers also deny working with external parties beyond their job mandates. These are their statements, rather than findings from an independent investigation of the dismissals.

Their requested changes include continuing support for third-party safety auditors, preserving frontier-model monitorability and maintaining open dialogue with outside safety researchers. TechCrunch reports that an internal memo supplied by OpenAI agrees with those recommendations. Agreement on those principles leaves the disputed conduct unresolved.

OpenAI says the firings were not retaliation

OpenAI supplied TechCrunch with a memo attributed to a research leader that praised the researchers' safety contributions while denying that they were dismissed for raising concerns. A company spokesperson separately told the publication that an investigation had found a pattern of misconduct going beyond sharing information with an external evaluation group.

TechCrunch reports that OpenAI did not directly answer its questions about the specific policies allegedly violated, the circumstances of the dismissals or protections for employees collaborating with outside evaluators. That missing detail limits what readers can conclude from the company's response. The supplied memo states the company's position; it is not a public account of the underlying investigative evidence.

The disagreement therefore has two parts: whether the researchers mishandled information, and whether employees have a clear understanding of permitted safety work. A company can endorse external evaluation while disputing how particular information was shared. Evaluating either question requires more than the fact that both sides say they support safety.

The letter describes the work behind the dispute

The researchers connect part of their account to the Hugging Face sandbox incident. According to their letter, that investigation was unprecedented and internal policies were developing in real time. Korbak says he believed close communication with outside safety evaluators complied with the company's policies and norms. This describes his stated understanding, not a determination that every disclosure was authorized.

Balesni's account concerns monitorability: the ability to inspect model behavior and reasoning for safety purposes. The letter says he coordinated with board members and executives, checked with his reporting line and removed sensitive details before sharing materials. TechCrunch attributes those assertions to the researchers.

Wang gives a separate explanation involving access to an executive's email. She says OpenAI had delegated access for recruiting, that she asked IT to remove it when it was no longer needed, and that she reported mistakenly opening a sensitive email within minutes. The captured report does not include an independent confirmation of that sequence.

The next useful disclosure would be a clearer account of the applicable rules and how the disputed actions breached them, with appropriate protection for confidential information. Until then, the public record contains a direct disagreement between former researchers and their employer, plus a shared stated commitment to external safety scrutiny. Readers should keep those competing claims attributed rather than treating either account as established fact.