OpenAI Cuts Ties With Three Safety Researchers Over Alleged Information Mishandling

Three Fired OpenAI Researchers Urge the Company to Preserve AI Monitoring and Outside Audits

Three fired OpenAI safety researchers have published a letter urging the company to preserve its ability to inspect AI reasoning—and the outside collaboration they say makes safety work possible. Tomek Korbak, Jasmine Wang, and Mikita Balesni warn that their dismissals risk silencing remaining staff. OpenAI disputes their account while supporting the letter’s safety recommendations.

The public letter addresses OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. Its requests reach beyond the authors’ own departures: preserve model monitorability, maintain independent auditor access, and spell out how employees can work with external safety organizations. The researchers argue that unclear boundaries now threaten both internal debate and third-party accountability.

Chain-of-thought monitoring examines the reasoning text a model generates before answering, looking for signs of unwanted behavior. In a December 18, 2025 research publication, OpenAI said this approach was substantially more effective than watching actions and final answers alone in almost every setting it studied. Monitorability describes how well a monitoring system can infer properties of an AI agent’s behavior from those signals.

The letter’s authors say that visibility is deteriorating in frontier models, the industry’s most advanced systems. They urge OpenAI and its competitors not to pursue developments that reduce it further while monitoring remains a safety tool. Their recommendation is a constraint on future development, not simply a request for better checks on a finished model. They argue the industry does not yet know how to safely develop and deploy systems it cannot monitor.

OpenAI’s earlier research also described monitorability as fragile to changes in training, data, and scale. But its experiments found that reinforcement learning at the scales studied did not materially degrade it. Those results covered two training runs, and OpenAI explicitly cautioned that larger runs could behave differently. That older finding does not resolve the researchers’ current warning; it shows why the particular training conditions matter when assessing whether oversight is holding up.

If conduct that was considered normal last month now constitutes grounds for sudden dismissal, everyone at OpenAI is left guessing where the line is.

The researchers say they acted within their job responsibilities and the working norms in place at the time. Their explanations concern different activities, rather than one shared act of disclosure:

Korbak says he was OpenAI’s technical contact for METR during the investigation of an incident involving Hugging Face. The letter says internal procedures were being developed in real time and close communication with external counterparts was essential. According to The Decoder, Korbak says he was verbally told his firing concerned communications with METR, without a written explanation.

Balesni says he coordinated cross-company commitments to preserve monitorability, with board members and senior executives involved in discussions. He says he checked with his reporting line and removed sensitive details before sharing materials. His account presents external communication as necessary to work that could not succeed within one company alone.

Wang says she had authorized access to an executive’s inbox for recruiting and asked IT to remove it when that work ended. According to the letter, IT did not revoke access, and messages appeared together without clearly distinguishing their destination inbox. She says she accidentally opened a sensitive email, reported it within minutes, and again requested removal.

OpenAI rejects the implication that the dismissals punished safety concerns. In its response, reported by The Decoder, the company said a thorough investigation found violations of clear sensitive-information policies and a significant breach of trust beyond the letter’s account. It did not publicly specify that additional conduct. The researchers also deny supplying a leak to The Information about allegedly less-monitorable architectures, saying that article undermined their cross-company work.

The letter asks OpenAI to honor what the authors describe as Sam Altman’s September 12 public commitment to give independent evaluators ongoing, employee-like access. They specifically worry that their firings could become a reason to end work with METR or sharply restrict auditors’ access and scope. That is their concern about a possible consequence, not an announced withdrawal by OpenAI.

OpenAI says it is working on contracts with outside safety auditors and agrees that preserving frontier-model monitorability requires an industry-wide commitment. Gizmodo also reported that an internal company memo strongly agreed with the letter’s recommendations. Agreement on those goals does not settle the employment dispute: OpenAI maintains that policies were violated, while the authors maintain that their work followed the applicable norms.