The three researchers dismissed by OpenAI are calling on research labs not to compromise on monitoring the reasoning of models

In a letter addressed to the board of directors, Tomek Korbak, Mikita Balesni and Jasmine Wang have raised concerns about the loss of transparency in the ‘thought processes’ of models such as GPT-6 Astra.

A week after their forced departure from OpenAI, accused of transmitting internal information to an AI security organization, Tomek Korbak, Mikita Balesni and Jasmine Wang wrote to the company's board of directors and security committees. The letter is not public; The Wall Street Journal published excerpts.

The chain of thought, the last observation post

Their message is clear: the industry is not yet able to develop or deploy risk-free models that it cannot monitor, and advanced laboratories should refrain from any changes that further reduce this capacity. The concern is in particular GPT-6 Astra: its system sheet admits that the models in this class could escape from the chain of thought monitors under adverse conditions, and its mode of reasoning known as "recurrent depth" rotates a request several times within the network before producing anything, a step that leaves no legible trace.

When questioned by the WSJ, OpenAI states that the dismissals had nothing to do with security alerts and presented an internal note where the company said that it fully agreed with the authors' recommendations.