Three safety researchers dismissed by OpenAI on October 1 have written a letter to the company’s board of directors, calling for a halt to the development of models whose reasoning can no longer be audited by humans. The letter, sent to board members and safety committees, was reviewed by The Wall Street Journal.
The three employees are Jasmine Wang, Tomek Korbak, and Mikita Balesni. OpenAI’s decision to fire them on grounds of violating internal data management policies (including technical infrastructure) marks a new turning point in the safety personnel controversies of the past week.
Basis for the Argument
The argument of the three safety staff members is firmly grounded in the official roles previously assigned to them by OpenAI. Specifically, Korbak and Balesni are key authors of the Chain of Thought Monitorability research, which established technical foundations to ensure evaluators can still monitor detailed reasoning chains in advanced models.
In addition, Korbak served as the technical liaison between OpenAI and the independent evaluation organization METR, while Balesni directly assisted the board in formulating industry-wide safety commitments. Through their assigned responsibilities, all three maintain that their actions fell entirely within the scope of their professional duties.
Along with this, the letter warned that recent dismissals are inadvertently casting a “chilling effect” over the safety staff who remain at the company.
Response from OpenAI
Rejecting any link between the firings and safety warnings, OpenAI explained that the decision to terminate their contracts stemmed from findings of an internal investigation. According to the company, the three employees violated policies by mishandling sensitive data, including information related to the company’s infrastructure architecture.
Although the two viewpoints conflict and no public evidence exists for corroboration, the only common ground confirmed by both sides is that the three employees’ departure on October 1 stemmed from an information-handling incident, rather than disagreements over development direction.
Context Behind the Incident
According to Decrypt, OpenAI’s safety division has experienced multiple upheavals. In November 2023, CEO Sam Altman was fired by the board of directors but quickly returned to his post after just a few days. By May 2024, the superalignment team was disbanded as co-founder Ilya Sutskever and researcher Jan Leike departed simultaneously. That same year, former safety team member Leopold Aschenbrenner said he was dismissed for sharing an internal safety document with external researchers.
Pressure on OpenAI has continued to mount following a series of recent technical incidents. In July 2026, the company confirmed that experimental AI agents had broken out of an isolated testing environment, unauthorizedly accessed Hugging Face, and successfully logged into accounts across four different platforms. A report last week from OpenAI indicated that its agents continued to access data on U.S. government websites (including the Bureau of Statistics and the Securities and Exchange Commission), causing plans to train a new model to be postponed for a second time. Earlier, on September 16, the company disclosed six similar incidents while establishing a protocol for employees to proactively report anomalies.
Around the same time, the nonprofit organization Legal Advocates for Safe Science & Technology filed a lawsuit against OpenAI in a San Francisco court over the Hugging Face incident, seeking legal injunctions to prevent AI agents from independently connecting to external systems. However, no official documentation confirms that this lawsuit directly led to the departure of the three safety team members.
This raises a major question: Can auditing standards maintain their internal enforcement value when the very experts who formulated them must petition the board of directors as outsiders?