The dispute began last week, when OpenAI dismissed safety researchers Jasmine Wang, Tomek Korbak and Mikita Balesni after an internal investigation. The company said it found a “pattern of misconduct” involving the handling of sensitive research information and maintained that the case went beyond contact with an outside AI evaluation group.¹

The researchers reject that account. In an open letter to OpenAI’s safety bodies, they said they had acted within the company’s mission and normal working practices — particularly during the unprecedented investigation into an agent breach involving Hugging Face. They argued that safety staff must be able to consult outside experts without fear: “The freedom to do so without fear, and to have well-defined internal procedures that enable this work, is itself an essential safety mechanism.”¹

Their individual explanations sharpened the conflict. Balesni said his work with third-party safety organizations had been coordinated with managers, research leadership and board members; Korbak said communication with the nonprofit METR was part of his job. Wang said she was dismissed over accidentally opening an executive email after IT failed to remove recruiting access she had asked to relinquish.²

OpenAI’s internal response, as described in reporting, struck a different note: a research leader praised the trio’s safety contributions and said the company had “always encouraged” employees to raise concerns. “We do not terminate employees for raising concerns,” the memo said.²

But the former employees say the lack of detailed written explanations leaves colleagues unable to tell which conduct is safe. They warn the dismissals could be used to curtail OpenAI’s work with METR, whose independent evaluations have become especially consequential amid scrutiny of rogue-agent incidents.³

The practical disagreement is now stark. OpenAI frames the episode as enforcement of information-handling rules; the researchers see it as a test of whether independent oversight and internal dissent can survive when safety work becomes uncomfortable. Their recommendations — embedded third-party auditors, monitorable frontier models and open dialogue with the wider safety community — are measures OpenAI’s memo reportedly said it supported.²