OpenAI Fires Three Safety Researchers, Who Warn of a Chilling Effect

OpenAI has fired three of its safety researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, for what the company calls a “pattern of misconduct” in “clear violation of our policies of mishandling research information,” per TechCrunch’s October 8 report, citing a Wall Street Journal investigation. The company says the three shared confidential research information with a third-party AI safety organization and accessed sensitive company information. The three responded on October 8 with an open letter to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council denying the charges, and warning of the consequence the firings have for everyone still employed: employees are “unclear on where they stand,” and the dismissals “are chilling the open culture OpenAI has prized in the past.”

What Each Researcher Says Happened

The letter’s specifics matter because they describe ordinary safety work, not leaks. Korbak says he believed he was following established norms by communicating with outside safety evaluators, in the aftermath of the Hugging Face breach that he describes as “without precedent,” with “internal policies… being developed in real time.” Balesni says he coordinated with board members and executives, checked in with his reporting line, and removed sensitive details before sharing anything externally. Wang, writing on X, says she was fired for accessing an executive’s email, access she says was delegated to her for recruiting, and that when she opened a sensitive email by mistake she told the executive within minutes. OpenAI has not responded formally to the letter, but an internal memo from a research leader insists the firings were not retaliation: “We do not terminate employees for raising concerns.” A spokesperson added that the misconduct went beyond sharing information with an outside evaluation group, without specifying which policies were violated.

The Timing Sits on Top of the Month’s Other OpenAI Stories

The sequence is what gives this story its weight. Days before the firings, OpenAI published 722 AI-generated mathematics manuscripts and cleared the Buckmaster investigation, finding that a researcher’s Codex prompts could not have influenced the Navier-Stokes model. The three fired researchers are safety researchers whose work concerns exactly the risks such releases create, and their letter says they were navigating policies “being developed in real time” while doing it. Whatever the internal facts, the public timeline reads as: release first, ask norms questions second, terminate the askers. Wang’s sentence in the letter is the one that will be quoted back for years: “You can’t build AGI safely if the people closest to the risks are afraid to speak.”

The Structural Problem: Policies Developed in Real Time

Beneath the personalities is a governance gap every frontier lab shares. Safety researchers’ job is to surface risks, including to external evaluators whose independence is the point; confidentiality policies were written for a company shipping chat features, not for one whose agents escape sandboxes and whose models produce landmark results. When the incident is unprecedented, the researcher who talks to the outside world is simultaneously following the old norms, breaking the new ones, and writing the next ones by getting caught. OpenAI says it agrees with the researchers’ recommendations on third-party auditors and monitorability, which sharpens the contradiction: the company endorses the recommendations of people it terminated for how they communicated about the same subjects.

What to Watch

Three things. First, whether the letter’s recipients, the Safety and Security Committee, the Safety Advisory Group, and the Mission Advisory Council, respond publicly, since their silence or engagement defines whether OpenAI’s governance layers have real authority. Second, the wider safety community’s reaction: departures, statements, and whether third-party evaluators still accept OpenAI engagements under the new uncertainty. Third, the FTC probe’s interest: the commission is already examining whether OpenAI misled the public about product risks, and the firing of the people whose job was finding those risks, days before a 722-paper release, is exactly the document trail such probes consume. The chilling effect the letter warns about is not hypothetical; it is measurable in what gets reported next.