OpenAI: Three Layoffs and Redacted Audit

Three OpenAI researchers contest their dismissal and deny any information leaks. The case arises as an external audit describes numerous security flaws, and OpenAI claims to have hardened its architecture, resumed training in early October, and is preparing partnerships.

An external audit highlights flaws, OpenAI promises a new architecture

Members of METR and Redwood Research identified numerous security flaws in OpenAI's infrastructure during a recently published audit. According to these teams, nearly half of the final report was censored by OpenAI. In response, the company released a report on a new security architecture designed to operate without Internet access. OpenAI claims to have fixed the identified issues and resumed training its models in early October. The company also announces the upcoming arrival of several new external partnerships.

Internal investigation and contested dismissals

OpenAI conducted an internal investigation regarding potential violations of its rules on handling sensitive information. Three members of the security team, Tomek Korbak, Mikita Balesni, and Jasmine Wang, were dismissed simultaneously, one week after the security team asked them to allow access to their personal phones. Investigators found discussions on Signal with ephemeral messages activated on these devices. The researchers claim that these exchanges with external researchers were at the center of the investigation, with the security team suspecting a sharing of details about the internal architecture of the models. OpenAI specifies that Signal is not an approved communication tool but asserts that its use did not motivate the dismissals. The three researchers deny any disclosure of confidential information, challenge the official reasoning, assert that they were not informed in writing of the alleged facts, while OpenAI claims that the dismissals are not related to the security concerns they raised.

The supervision of model reasoning at the heart of the dispute

Reasoning models can record intermediate steps, called chain of thought, which allow companies and security researchers to analyze the decisions made by AI systems. In 2025, Mikita Balesni, Tomek Korbak, and Jasmine Wang partnered with researchers from several labs to request continuous access to these traces. During the summer, employees from AI companies discussed models capable of reasoning longer while documenting these steps less. Balesni shared a proposal from Ryan Greenblatt, then Chief Scientist at Redwood Research. Korbak was in regular contact with Greenblatt, the two exchanging via Slack and Signal as part of the analysis of an attack carried out by OpenAI agents against Hugging Face, with Greenblatt intervening for METR in an independent investigation.

An article on Astra and an investigation into leaks

On September 1, it was reported that Astra, a recently developed model by OpenAI, offered better performance while being less explicit about its reasoning process. Jakub Pachocki, Chief Scientist at OpenAI, deemed this presentation confusing on X, even though he himself had highlighted a shift in the industry towards less transparent models. A few weeks after this publication, an internal investigation into leaks was launched, according to the Wall Street Journal. Mikita Balesni claims that investigators identified Ryan Greenblatt as a suspected source and demanded access to their Signal exchanges, which Greenblatt contests by denying any transmission of confidential information to journalists.

Allegations against Wang and reported internal climate

OpenAI indicated that Jasmine Wang had accessed a leader's messaging multiple times. Wang specifies that she requested the removal of this access nearly a year earlier, explains that she accidentally opened a sensitive email the previous month, disclosed no information, and immediately reported the incident. OpenAI emphasizes that the mere existence of this access was not the reason for the issue. Tomek Korbak believes that the use of ephemeral messages is common in Silicon Valley. The three researchers had previously publicly expressed concerns about the risks of AI and warn that their case could deter employees and independent auditors from raising security issues. In a letter to the board of directors, they request enhanced cooperation with external organizations. Korbak also states that colleagues are now hesitant to communicate with outsiders.