Artificial Intelligence
Fired OpenAI researchers dispute leak allegations
Three former OpenAI safety researchers denied leaking confidential material, said they were not told in writing what they had supposedly mishandled, and linked the dispute to their work on model transparency.
- OpenAI dismissed Tomek Korbak, Mikita Balesni and Jasmine Wang after an internal investigation into alleged policy breaches involving sensitive information.
- The three told the Wall Street Journal that they denied sharing confidential information and said they were never told in writing what they were said to have mishandled.
- OpenAI said the firings were not tied to the safety concerns the researchers had raised.
- The company’s security team suspected information about internal model architecture had been shared with an external safety researcher and searched the researchers’ phones, where it found Signal chats with disappearing messages enabled.
- Balesni said investigators named Ryan Greenblatt as the suspected source, while Greenblatt denied giving confidential information to journalists.
OpenAI dismissed Tomek Korbak, Mikita Balesni and Jasmine Wang after an internal inquiry into alleged handling of sensitive material. In interviews with the Wall Street Journal, the three said they denied leaking confidential information and believed their safety advocacy may have played a part in their removal.
OpenAI told the Journal that the dismissals were not about the safety issues the researchers had raised. The company’s security team had suspected that details about its internal model architecture had been shared with an outside safety researcher, and it asked to inspect the researchers’ phones.
On those devices, the team found Signal conversations with disappearing messages turned on. OpenAI said Signal was not an approved tool for company use, although it said that was not the reason for the terminations.
The dispute also centred on how AI models’ reasoning can be examined. The three researchers had joined a 2025 paper with other lab researchers calling for continued access to chain-of-thought traces, and Balesni had shared a related proposal from Ryan Greenblatt.
Korbak said the case was making colleagues less willing to speak with outsiders or raise concerns internally. Separately, the story said a later audit by METR and Redwood Research found security gaps in OpenAI’s infrastructure, while OpenAI redacted nearly half of that report and said it had fixed the issues and resumed training in early October.
Named in this story
People
- Tomek Korbak
- one of the fired safety researchers who disputed the leak claims
- Mikita Balesni
- one of the fired safety researchers who disputed the leak claims
- Jasmine Wang
- one of the fired safety researchers who disputed the leak claims
- Ryan Greenblatt
- researcher named by investigators as a suspected source
- Jakub Pachocki
- OpenAI chief scientist who disputed The Information’s report
Companies
- OpenAI
- dismissed the three researchers after an internal leak inquiry
- Hugging Face
- site of the incident Greenblatt analysed with Korbak
Organisations
- Wall Street Journal
- interviewed the researchers and relayed OpenAI’s account
- Redwood Research
- research group linked to Greenblatt and to safety work cited in the story
- METR
- hired Greenblatt for an independent investigation and joined an audit of OpenAI
- The Information
- published the report on Astra that preceded the leak investigation
Products and systems
- Signal
- messaging app used in chats examined during the inquiry
- Astra
- OpenAI model reported to reveal less of its reasoning process
How the source tells it
The article is mostly straight reporting, with brief urgency around the consequences of the dispute over AI transparency and internal safety work.
- urgency frames the transparency dispute as having major consequences and as becoming more pressing over time