Fired OpenAI safety researchers: never told in writing what they allegedly leaked
Per THE DECODER’s 2026-10-24 follow-up attributing Wall Street Journal interviews, Tomek Korbak, Mikita Balesni, and Jasmine Wang deny leaking confidential information and say they were never told in writing what they supposedly mishandled. OpenAI says the firings were unrelated to raised safety concerns and that Signal was not the reason. Background includes chain-of-thought monitorability advocacy and Signal disappearing messages with Redwood’s Ryan Greenblatt.
THE DECODER on 2026-10-24 followed up on three OpenAI safety researchers—Tomek Korbak, Mikita Balesni, and Jasmine Wang—fired after an internal probe into alleged mishandling of sensitive information. In Wall Street Journal interviews summarized by Decoder, they deny sharing confidential information, question the official rationale, and suspect a link to their advocacy for more transparent systems. Decoder notes OpenAI told the Journal the firings had nothing to do with safety concerns the researchers had raised, while the three say they were never told in writing which information they supposedly leaked and worry the case could chill outside auditors. This piece dates the follow-up coverage, not the original termination day.
As Decoder retells the Journal’s account, OpenAI’s security team suspected the three of sharing internal model-architecture details with an outside safety researcher and asked them to unlock personal phones; investigators found Signal chats with disappearing messages. All three were fired about a week later. OpenAI said Signal was not an approved tool but insisted it was not the reason for termination; Korbak called disappearing messages routine Silicon Valley practice. A technical flashpoint is whether outsiders should keep inspecting models’ chain-of-thought traces. Balesni, Korbak, and Wang joined a 2025 paper urging continued access; when staff debated models that think longer while writing fewer intermediate steps, Balesni shared a related proposal from Ryan Greenblatt, then chief scientist at Redwood Research.
Korbak also said he stayed in contact with Greenblatt over Slack and Signal while they worked on the analysis of OpenAI agents’ attack on Hugging Face, for which Greenblatt had been hired by METR. Decoder writes that after The Information’s September 1 report that Astra was more capable but revealed less reasoning, OpenAI opened a leak probe; Balesni says investigators named Greenblatt as a suspected source and demanded Signal chats, which Greenblatt denies using to tip journalists. Separately, OpenAI accused Wang of repeatedly accessing an executive’s inbox; she says she had asked nearly a year earlier to revoke that access, accidentally opened a sensitive email the prior month, did not share it, and reported it immediately—OpenAI saying the access itself was not the issue. The three wrote the board urging closer work with external safety orgs. All of the above is Decoder’s attribution of competing WSJ/party/company claims.