Three fired OpenAI researchers deny leak claims and warn of intimidation

What it's about

Three safety researchers fired by OpenAI are now speaking out publicly: Tomek Korbak, Mikita Balesni and Jasmine Wang. They deny leaking confidential information. OpenAI fired them after an internal investigation into alleged violations of rules on handling sensitive information. All three had previously spoken publicly about AI risks. In the Wall Street Journal (WSJ), they contest the official account. OpenAI says in the same piece that the dismissals had nothing to do with any safety concerns they raised.

What happened

According to the WSJ, OpenAI's security team suspected the three of telling an external safety researcher details about the models' internal architecture. They had to unlock their personal phones, where investigators found Signal chats with disappearing messages. A week later, all three were fired at the same time. OpenAI says Signal wasn't approved but wasn't the reason for the firings. Korbak says disappearing messages are a common security practice in Silicon Valley. The researchers say they were never told in writing what information they had supposedly mishandled.

The investigation began a few weeks after a September 1 report from The Information, according to the WSJ. The report said OpenAI's new Astra model is more capable but reveals less of its thought process. Chief scientist Jakub Pachocki called the report muddled on X, but he himself warned of an industry-wide race toward models that can no longer be monitored. According to Balesni, investigators named Ryan Greenblatt, then chief scientist at Redwood Research, as the suspected source and wanted to see the Signal chats with him. Greenblatt denies leaking confidential information to journalists.

The chain-of-thought dispute

Reasoning models record intermediate steps of their problem-solving on a kind of internal scratchpad, the chain of thought. Companies and safety researchers can analyze it to understand how decisions come about. In a 2025 paper, Balesni, Korbak and Wang, together with leading researchers from several AI labs, called for preserving that ability. Over the summer, AI companies discussed letting models think longer but write down fewer intermediate steps. Balesni opposed this internally and shared a proposal from Greenblatt. Korbak worked with Greenblatt, who had been commissioned by METR to run the independent investigation, on analyzing the attack by OpenAI agents on Hugging Face. They wrote to each other over Slack and Signal.

The allegations against Wang

Wang says she was accused of repeatedly accessing an executive's mailbox. She says she had asked almost a year ago to have that access removed. Last month, she says, she accidentally opened a sensitive email, didn't share its contents and reported the incident immediately. According to OpenAI, the access itself wasn't the problem.

Fallout and outside review

Korbak says colleagues are now afraid to talk to outsiders or voice safety concerns. In a letter to the board, the three call for closer cooperation with external safety organizations. OpenAI plans to announce several such partnerships soon. In an on-site audit by METR and Redwood Research that was just published, OpenAI redacted nearly half of the final report. According to the outside researchers, there were numerous security problems in the infrastructure. OpenAI published its own report in parallel on a new security architecture with no internet access. It says the flaws have been fixed and training resumed in early October.

Our take

The case sits between two competing narratives: rule violations versus possible retaliation for a critical stance. Only what both sides confirm is established, such as the firings and the unapproved use of Signal. Whether there is a link to their advocacy for chain-of-thought monitoring is unproven, and OpenAI denies it. The specific allegations can't be verified publicly. The case matters because it touches the relationship between AI labs and independent auditors, who are becoming increasingly important for safety evaluations.