OpenAI fires three safety researchers after leak inquiry into model design

OpenAI dismissed three safety researchers after an internal investigation into the possible leak of confidential model design information, The Wall Street Journal reported on October 22, 2026. The three were Tomek Korbak, Mikita Balesni and Jasmine Wang.

The investigation examined Signal messages the researchers exchanged with Ryan Greenblatt, an outside AI safety researcher. According to Balesni, investigators suspected Greenblatt, then chief scientist at Redwood Research, of being the source of the leak. OpenAI's security investigation team asked employees, including the three researchers, to unlock their personal phones. The checks showed that some researchers had disappearing-message settings enabled in conversations with Greenblatt.

A week later, the three were told of their dismissal in separate meetings held at the same time. Korbak and Balesni said management cited the use of Signal's auto-deleting messages as a reason. Wang said she was told she had repeatedly opened an executive's email; she said she had asked about a year earlier to have that access removed.

OpenAI said the three violated clear rules on handling sensitive information and were not fired for raising safety concerns. It added that Signal use was not authorized but was not the only reason for the dismissals. Whether the three actually leaked confidential information has not been confirmed.

Concerns over Recurrent Depth

The investigation concerned design information around Recurrent Depth, also called loop transformer, which reuses the same neural network layers repeatedly. Safety researchers are concerned that much of its internal computation may not appear as natural-language chain-of-thought, making monitoring harder. In early September 2026, The Information reported on ASTRA's new inference technology, and OpenAI chief scientist Yakub Pachoński criticized the report as confusing while warning against racing to build models the industry cannot monitor.