OpenAI Safety Researchers Fired Over Data Dispute
OpenAI fired three senior alignment researchers for allegedly leaking data, while the team claims retaliation over model safety warnings.
Dismissal of scientists working on reasoning transparency sparks concerns over third-party safety audits.
On October 1, 2026, The Wall Street Journal reported that OpenAI fired three senior alignment researchers: Tomek Korbak, Mikita Balesni, and Jasmine Wang. The company confirmed the dismissals, citing internal policy violations after the researchers shared materials with METR (Model Evaluation and Threat Research), an independent nonprofit that evaluates risks in frontier AI models. The incident puts OpenAI safety researchers at the center of an escalating dispute over oversight.
Why OpenAI Safety Researchers Track Reasoning
Korbak and Balesni led a widely cited study published on arXiv examining chain-of-thought monitorability. In advanced artificial intelligence models, chain of thought works like scratch paper during a difficult exam: before delivering a final answer, the system writes down intermediate steps in plain text. That visibility allows human reviewers and automated tools to verify whether a model is bypassing rules, hiding intentions, or executing unsafe code.
Their work reached beyond internal laboratory experiments. Korbak served as OpenAI’s official technical liaison to METR, while Balesni advised the company’s board of directors on safety protocols. Because of those responsibilities, the dismissals drew immediate attention across the AI research community.
Conflicting Accounts and the RUMOR Label
OpenAI’s official statement clashes directly with the account provided by the researchers. In comments to the press, the company maintained that an internal inquiry uncovered a breach of trust and unauthorized data sharing. Management has not publicly identified which specific files or benchmarks were shared with METR.
In response, the three researchers issued an open letter sent to board committees. They stated that they were dismissed after repeatedly warning leadership against deploying opaque models that shield reasoning from human inspection. The letter also described an atmosphere of fear and self-censorship among remaining safety staff. Without an independent forensic audit of internal communications, the allegation of direct retaliation remains classified as a RUMOR by our newsroom, as does the exact content of the shared files.
Release Schedules and the Industry Precedent
The dismissals surfaced two days after The New York Times reported that OpenAI executives had brushed aside internal warnings to keep commercial product rollouts on schedule. The departures also mirror the 2024 dismissals of Leopold Aschenbrenner and Pavel Izmailov, whom OpenAI fired under similar accusations of leaking confidential data.
For the broader tech ecosystem, the dispute exposes structural vulnerabilities in third-party oversight. If developers hide internal reasoning steps to lower compute costs or protect commercial secrets, external evaluators lose their primary window into model behavior. Treating technical collaboration with testing organizations as a disciplinary breach risks eroding collective safety baselines just as autonomous systems take on wider real-world roles.