OpenAI fires three safety researchers over alleged leaks

The internal investigation

OpenAI recently dismissed three safety researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, following an internal probe into policy violations. The company stated the investigation focused on how the employees handled sensitive information. However, the researchers have publicly disputed the company's account of the events.

According to reporting from The Decoder, the security team suspected the three were sharing internal model architecture details with an outside researcher. This led to a request for the employees to unlock their personal phones. Investigators discovered chats on the Signal app that utilized a disappearing messages feature.

While OpenAI confirmed that Signal was not an approved communication tool, they insisted it was not the reason for the firings. Korbak countered this by stating that disappearing messages are a common security practice in Silicon Valley. The researchers maintain they never shared confidential data and were never told in writing what specific information they allegedly leaked.

The transparency dispute

At the heart of the conflict is a technical debate over how much of an AI's internal logic should be visible to the public. Reasoning models can record intermediate steps of their problem-solving process on an internal scratchpad called a chain of thought. Safety researchers rely on these traces to understand how a model arrives at a specific decision.

The three fired researchers had recently co-authored a paper calling for continued access to these reasoning steps. They pushed back internally against a trend toward models that think longer but write down fewer intermediate steps. This cooperation became a point of friction after reports suggested OpenAI's new model, Astra, revealed less of its reasoning process than previous versions.

The researchers were also in contact with Ryan Greenblatt, a former chief scientist at Redwood Research. Greenblatt was hired by METR to conduct an independent investigation into an attack by OpenAI agents on Hugging Face. The investigation into these communications eventually led to the internal leak probe that resulted in the three terminations.

Accusations and defenses

The investigation into the three employees involved different specific allegations. Balesni reported that investigators named Greenblatt as a suspected source and demanded to see his Signal chats. Greenblatt has denied passing any confidential information to journalists during the process.

Jasmine Wang faced a separate accusation regarding her access to an executive's inbox. She explained that she had requested that access be revoked nearly a year prior and only opened a sensitive email by accident the month before the incident. She claims she reported the accidental opening immediately and did not share the contents with anyone.

OpenAI maintains that the mere existence of the access was not the issue for the termination. The researchers, however, argue that the timing of the firings suggests a connection to their public advocacy. They expressed concern that the move sends a message to other employees about the risks of raising safety concerns or speaking with outsiders.

Impact on AI oversight

The fallout from these firings extends beyond the individuals involved and into the broader AI safety community. Korbak noted that colleagues are now afraid to talk to external parties or voice concerns internally. The researchers sent a letter to OpenAI's board calling for closer cooperation with external safety organizations to prevent this type of friction.

This situation occurs alongside a recent on-site audit of OpenAI's infrastructure. Staff from METR and Redwood Research identified several security gaps during the review, but OpenAI redacted nearly half of the final report. The company has since published its own report on a new security architecture that operates without internet access.

OpenAI claims to have fixed the identified issues and resumed training in early October. While the company plans to announce new external partnerships soon, the incident highlights the difficulty of balancing proprietary secrets with the need for independent verification. The researchers warn that this case could discourage auditors from working with AI companies altogether.

What comes next

The industry will be watching to see if OpenAI follows through on its promise of new external partnerships. These relationships will be critical in determining whether independent researchers can continue to audit model reasoning. The tension between

The future of audits

The industry will be watching to see if OpenAI follows through on its promise of new external partnerships. These relationships will be critical in determining whether independent researchers can continue to audit model reasoning. The tension between corporate security and public safety remains a central theme for the next generation of models.

If OpenAI moves toward more transparent cooperation, it may set a standard for how other labs handle external audits. Conversely, if the current friction continues, it may lead to a more closed environment where only internal teams can see how models think. The outcome of this specific dispute will likely serve as a bellwether for the future of AI governance.