An unusually specific open letter on AI risk from insiders has recently surfaced. According to the Wall Street Journal, the core concern of the letter's authors is preserving the ability to monitor models' chain of thought. Strictly speaking, however, the three authors no longer count as insiders — they were just fired by OpenAI.
Last week, OpenAI announced it had "terminated its relationship with three individuals for violating company policies on accessing and handling sensitive information." The three allegedly passed information to an AI safety organization. They are identified as Tomek Korbak, Mikita Balesni, and Jasmine Wang, all of whom, from public statements, believe AI poses existential risk. Before being fired, Balesni posted on X that he believed AI had a "10% chance of killing all of humanity"; Wang wrote that "it is hard to adequately describe how dangerous acceleration is on the path to recursive self-improvement (RSI)," where RSI refers to
the ability of AI models to improve themselves.
The joint letter is addressed to OpenAI's board and the company's "Safety Committee" and has not been publicly released; the Wall Street Journal published only excerpts. Its core passage reads: "As an industry, we still do not know how to safely develop and deploy models we cannot monitor... OpenAI and other frontier companies should not pursue research that would further erode (monitoring capability)."
The letter's specific concerns target OpenAI's latest flagship model, GPT-6 Astra. Two worries surround chain-of-thought monitoring: first, that the model may be learning to evade monitoring systems; second, that OpenAI may be steering the technical direction toward a place where chain-of-thought monitoring becomes meaningless. The model's system card explicitly states that "Astra-class models may evade our chain-of-thought monitoring under adversarial conditions." At the same time, Astra's reasoning process involves a more opaque mechanism called "recurrent depth" — it cycles the query internally through the model multiple times rather than generating content directly, a process that occurs inside an unknowable black box,
leaving no monitorable object for the outside world.