Technology & Science
OpenAI safety researchers dispute firings as outside audits come into focus
The researchers and OpenAI now give sharply different accounts of three dismissals. Their public dispute also tests how the lab will put its stated commitments to outside safety assessments into practice.
Executive summary
- Two accounts of the dismissals: OpenAI said an investigation found three safety researchers violated policies for handling sensitive information and that the firings were not about raising concerns. In a letter to the company’s safety oversight groups, the researchers said they believed they had acted within the working norms and mandates in place at the time. Public records reviewed do not independently resolve the disputed conduct or the company’s investigation.[1][2][3]
- Outside evaluation is the policy question: The researchers urged OpenAI to keep working with external safety groups, preserve the ability to monitor frontier models and clarify rules for employees’ collaboration with outside experts. OpenAI had publicly committed in September to supporting independent assessments with deep levels of access; the dispute does not establish whether those commitments have changed.[2][4]
- Evidence remains incomplete: OpenAI has not publicly identified the specific policies or conduct underlying the dismissals in the reporting reviewed. The researchers’ letter is their own account, while the company disputes their interpretation. Neither document supplies a neutral finding on the employment dispute.[1][3][2]
OpenAI said Friday that it fired three safety researchers after an investigation found violations of its policies for handling sensitive information. The company said the dismissals were not about raising safety concerns or speaking out, the Associated Press reported. Its public statement described a breach of trust but did not identify the specific information or policies at issue.[1]
The three researchers — Tomek Korbak, Jasmine Wang and Mikita Balesni — wrote to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council that they believed they had acted within their jobs and the company’s norms at the time. They said the investigation into a prior incident involving OpenAI agents and Hugging Face unfolded while internal procedures were being developed, and that unclear boundaries for outside collaboration had left employees uncertain. These claims are the researchers’ account; the company disputes their characterization of the firings.[2][1]
The researchers also denied being the source of a separate leak to The Information about model monitorability. OpenAI’s statement, as reported by AP, said its investigation found policy violations; the reporting reviewed did not provide an independent adjudication of the underlying events.[2][1]
The researchers asked the company to continue embedding third-party safety auditors, preserve monitorability of frontier models and set clear rules for collaboration with external safety organizations. OpenAI’s September policy says independent assessments should receive deep access across training, evaluation and deployment, and identifies monitoring safeguards and critical incidents among its proposed assessment areas.[2][4]
OpenAI told AP it agreed with the recommendations, referring the news agency to its public statement. That establishes a stated policy position, not whether the researchers’ departures will affect particular external assessment arrangements. The letter warns that they could; no change to an audit or monitoring program was documented in the sources reviewed.[1][2][4]
TechCrunch reported that OpenAI described a broader pattern of misconduct and said the case involved more than one exchange with an outside evaluation group. The company did not detail the alleged conduct to the outlet. The researchers, for their part, described work with outside evaluators as part of their roles and denied a separate leak. Without specific findings or records from the investigation, the competing public accounts leave key facts unresolved.[3][2]
The immediate governance question is whether OpenAI will turn its stated principles for independent assessment into stable working arrangements and clear employee procedures. The public materials set out the company’s commitments and the researchers’ request, but they do not show how those arrangements will operate after the dispute.[4][2]