Fired OpenAI Safety Researchers Say They Were Never Shown Proof of Leaks

Three safety researchers dismissed by OpenAI over alleged leaks are publicly disputing the company's claims, asserting that management never specified in writing what confidential details they supposedly disclosed.

In interviews with the Wall Street Journal, Tomek Korbak, Mikita Balesni, and Jasmine Wang denied sharing proprietary information. Instead, they suspect their abrupt terminations were connected to their internal advocacy for greater model transparency—specifically regarding whether external auditors should be able to inspect how reasoning models solve problems.

The Fight Over Chain-of-Thought Transparency

At the heart of the researchers' advocacy was an intensifying industry debate: should AI developers allow third-party researchers to observe intermediate reasoning steps?

Reasoning models frequently utilize an internal scratchpad—known as a chain of thought—to log intermediate logic while solving complex tasks. Both internal teams and independent safety auditors depend on these traces to evaluate whether models arrive at conclusions safely. In a 2025 paper, Balesni, Korbak, and Wang joined prominent researchers across several AI laboratories in calling for continued external access to this diagnostic capability.

The issue escalated over the summer as labs explored models that deliberate for longer periods while committing fewer intermediate reasoning steps to the scratchpad. Balesni pushed back against this trend internally, circulating a proposal authored by Ryan Greenblatt, then chief scientist at Redwood Research.

Korbak had also been collaborating with Greenblatt over Slack and Signal on an independent investigation run by METR into a high-profile incident involving OpenAI agents attacking Hugging Face.

A Media Leak and Phone Searches

Tensions flared following a September 1 report by The Information, which revealed that OpenAI's upcoming model, Astra, displayed higher capabilities but exposed fewer of its internal reasoning steps. Although OpenAI chief scientist Jakub Pachocki had previously warned about the risks of building models that cannot be monitored, he dismissed the report publicly on X as muddled.

Shortly after the article appeared, OpenAI launched an internal leak investigation. Security investigators identified Greenblatt as the suspected recipient of confidential leaks and demanded that Balesni, Korbak, and Wang unlock their personal mobile devices.

During the search, investigators discovered Signal messaging threads with disappearing messages activated. All three researchers were dismissed one week later. While OpenAI pointed out that Signal was not an authorized workplace communication platform, it insisted that the use of Signal itself was not the reason for the firings. Korbak countered that disappearing messages represent standard operational security among Silicon Valley researchers.

Wang faced an additional inquiry regarding recurring access to an executive’s inbox. According to Wang, she had requested that her access credentials be removed almost a year earlier, had accidentally opened a sensitive email the prior month, shared nothing, and self-reported the incident immediately. OpenAI told the Wall Street Journal that the inbox access was not the cause of the dismissal.

Impact on AI Safety Culture

OpenAI stated that the terminations were tied strictly to policy violations regarding sensitive materials and were unrelated to safety concerns raised by the trio. However, the three researchers maintain they were never given documentation identifying the sensitive information they allegedly breached.

Korbak warned that the firings have created a chilling atmosphere within OpenAI, leaving remaining staff reluctant to raise safety criticisms or consult external experts. In response, the three researchers submitted a formal letter to OpenAI’s board of directors urging closer structural partnerships with independent safety organizations.

The controversy coincides with wider transparency questions facing the lab. In a recent on-site infrastructure audit conducted by METR and Redwood Research, evaluators identified multiple security weaknesses; OpenAI subsequently redacted nearly half of the final audit report before publication. OpenAI later detailed a new isolated security architecture built without internet access, stating that training resumed in early October and promising further external safety partnerships soon.