OpenAI fires three researchers after phone inspection for Signal chats

The internal leak investigation

OpenAI recently conducted an internal investigation into a potential leak of sensitive information. The probe followed a report by The Information regarding Astra, OpenAI's latest model, and its specific techniques for processing information. The report suggested the model showed less of its internal thinking while becoming more powerful.

During this investigation, OpenAI's security team asked three safety researchers to open their personal phones. The goal was to identify any interactions with third parties through unsanctioned channels. Investigators specifically named Ryan Greenblatt as a suspected leaker, though he has denied these claims.

The company eventually fired the three researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, in a series of simultaneous meetings. OpenAI stated the firings were due to broken policies regarding sensitive information. A spokeswoman clarified that while using the Signal app was not approved, it was not the only factor in their dismissal.

Security practices vs. company policy

A central point of contention involves the use of the Signal messaging app. Security officers observed that some of the researchers had enabled disappearing messages in their chats with Greenblatt. The company cited this as a breach of trust during the leak inquiry.

The fired researchers argue that their actions were standard for the industry. Korbak stated that using disappearing messages is a basic security best practice in Silicon Valley. He suggested that the company may have reacted poorly to the researchers' high level of caution regarding security.

The researchers maintain that they never shared any confidential information outside of the company. They claim the company's reaction was disproportionate to their actions. Korbak suggested that the researchers may have been fired because they were trying to do too much on safety and security at once.

The role of Ryan Greenblatt

Ryan Greenblatt, who was a chief scientist at Redwood Research and now works at METR, is at the center of the leak allegations. He was contracted to examine logs from a previous incident where OpenAI's agents went rogue. He denies that he ever shared confidential information with reporters.

Greenblatt responded to the reports by stating that OpenAI made false accusations against him. He specifically denied allegations that he helped leak architecture details or read internal memos containing sensitive information. He claims he only saw versions of documents that were approved for sharing.

Greenblatt noted that his Signal chats were with professional contacts and that the app is common in the tech world. He believes OpenAI may have confused two different documents during their investigation. He maintains that he takes confidentiality seriously and has not shared internal data with the press.

Advocacy for AI reasoning

The three researchers had been actively pushing for more transparency regarding how AI models work. Specifically, they focused on the "chain-of-thought," which is the written scratchpad a model uses to solve problems. They wanted to ensure that this reasoning remained visible to researchers and the public.

In an October letter to the board, the researchers urged OpenAI to collaborate with outside safety groups. They wanted to ensure that the development of powerful models was balanced with rigorous oversight. This advocacy appears to have been part of the friction between the researchers and the company leadership.

The researchers believe their focus on safety and security may have contributed to their dismissal. Korbak suggested that the company might have felt the researchers were moving too fast or being too demanding. Despite the firings, the researchers continue to advocate for the importance of visible AI reasoning and safety protocols.

The impact on safety research

This situation highlights the difficult balance between corporate secrecy and the need for independent safety audits. By firing researchers who were pushing for outside collaboration, OpenAI may face challenges in maintaining a transparent safety culture. The incident underscores how quickly security protocols can be used to discipline employees.

The researchers' claims about being fired for