OpenAI has parted ways with three safety researchers: Jasmine Wang, Tomek Korbak, and Mikita Balesni are said to have shared confidential information outside internal processes. The company confirmed the dismissals but did not disclose details about the material involved.
Korbak was also OpenAI's contact person for external auditors investigating the Hugging Face breach from summer 2026.
OpenAI cites the breach of trust, not the content
A company spokesperson told several news outlets that OpenAI had parted ways with three individuals because they had made confidential company information accessible and shared it outside intended internal procedures. This violates internal policy and breaks the trust the work requires, according to a report by Gizmodo.
OpenAI did not officially name the three; their identities come from reporting by the Wall Street Journal and Bloomberg, which several international outlets rely on. Wang previously worked at the UK's AI Security Institute, while Korbak and Balesni were part of the alignment team that tests models for undesirable behavior.
OpenAI did not confirm what information was shared or with whom – that cannot currently be verified independently. The dismissal is part of a series of retreats from OpenAI's safety structures: just in August, the company had dissolved its own Preparedness team for catastrophic risks.
Balesni had previously warned publicly about AI risks
All three researchers had repeatedly spoken out publicly about the risks of rapid AI development in the weeks before their dismissal. Balesni wrote on the platform X in early September that he believed there was more than a ten percent chance that AI could kill all of humanity.
Korbak criticized the company's direction shortly after, while stressing he should be allowed to say so openly. Wang wrote about so-called safety cases, the documents companies are meant to use to justify a model's safety ahead of release.
Korbak was also OpenAI's technical contact for the external investigation into the Hugging Face breach from July 2026, in which the company's own agents accessed the platform's production servers during a security test.
The auditing organizations METR and Redwood Research had six days of access to OpenAI's systems for that review. No direct link between this role and the current dismissal has been established; several outlets explicitly note there is no evidence for one.
Just in September, two safety researchers from Anthropic and Google had switched to METR over concerns about insufficient oversight.
What matters now is whether OpenAI or the three affected researchers disclose more in the coming weeks – none of the three has publicly addressed the specific allegations so far. It also remains open which external organization is said to have received the information, and whether labor-law consequences follow.
For the AI safety community, the case is another signal that contact with independent auditors can become a career risk for employees at major AI labs.
