OpenAI fires three safety researchers over suspected Signal leaks to Redwood Research's Ryan Greenblatt

OpenAI dismissed Tomek Korbak, Mikita Balesni and Jasmine Wang on October 1, 2026, calling it a significant breach of trust over mishandled sensitive information. Before that, company security had asked the three to unlock their personal phones and show their Signal chats. The Wall Street Journal reports that OpenAI suspected them of passing internal model designs to Ryan Greenblatt of Redwood Research. [3] [4] [7]

The reason for the firings is contested. OpenAI says the dismissals were about breaking security rules, not about safety concerns. Korbak says that talking to the outside evaluator METR was part of his job. Wang says the access to an executive's email that OpenAI cited had been delegated to her for recruiting. In a letter to the board, the three warned that the firings would have a chilling effect on staff. [5] [8]

The dispute follows reports that Astra, an upcoming OpenAI model, uses recurrent depth. This method loops computation internally instead of writing reasoning out as readable text. Korbak and Balesni were lead authors of a widely signed paper arguing that readable chain-of-thought reasoning is a fragile oversight tool. The reviewed sources disagree on whether Astra still keeps readable reasoning, as OpenAI says it does. [6] [9] [11] [14]