Three security researchers recently dismissed from OpenAI have come forward to refute the company’s official explanation for their termination. Tomek Korbak, Mikita Balesni, and Jasmine Wang have denied leaking confidential information, questioning the stated reasons for their dismissal and suggesting their advocacy for increased AI transparency may have been the real cause.
OpenAI, following an internal investigation, concluded that the trio violated policies regarding sensitive information management, leading to their termination. Prior to their dismissal, all three had publicly expressed views on AI risks.
In interviews with The Wall Street Journal, the researchers countered the company’s narrative. OpenAI informed the newspaper that the dismissals were unrelated to the safety concerns they had raised. However, the researchers stated that the company never provided written specifics on what information was mishandled. They also expressed concern that this incident could deter independent safety auditors from collaborating with AI companies.
The Wall Street Journal report outlined the origins of the conflict. OpenAI’s security team suspected the three researchers of sharing internal model architecture details with an external security researcher, leading to a demand that they unlock their personal phones. Investigators reportedly found Signal chat logs with disappearing message features on the devices. A week later, all three were terminated. OpenAI acknowledged that Signal is not a company-approved communication tool but emphasized it was not the basis for the firings. Korbak, however, argued that disappearing messages are a common security practice in Silicon Valley.
The Core of the Dispute: Auditing AI’s Reasoning
At the heart of this dispute lies a significant technical issue: whether AI companies should permit external scrutiny of their models’ thought processes. Reasoning models record intermediate steps in an internal ‘chain of thought’ buffer, which companies and security researchers rely on to understand the logic behind a model’s decisions.
In a 2025 paper, Balesni, Korbak, and Wang, alongside leading researchers from various AI labs, called for the continued preservation of this auditable capability. This issue became more pressing over the summer as discussions began within multiple AI companies about a new class of models that might take longer to think but record fewer intermediate steps. Balesni reportedly raised objections internally, sharing a related proposal from Ryan Greenblatt, Chief Scientist at Redwood Research.
Korbak also maintained frequent communication with Greenblatt. The two collaborated via Slack and Signal, analyzing a high-profile incident where an OpenAI agent attacked Hugging Face. Greenblatt was commissioned by METR to lead this independent investigation.
The ‘Astra’ Report Ignites Internal Investigation
On September 1, The Information published a report suggesting OpenAI’s new model, Astra, was more capable but revealed less of its reasoning process. This report appears to have been the catalyst for the current trouble. At the time, OpenAI Chief Scientist Jakub Pachocki commented on X (formerly Twitter) that the report’s logic was muddled, despite having previously warned that the industry was racing to develop models that are difficult to monitor.
According to The Wall Street Journal, OpenAI initiated its internal leak investigation a few weeks later. Balesni stated that investigators identified Greenblatt as the recipient of leaked information and requested access to their Signal chat logs. Greenblatt has denied passing confidential information to journalists.
Accusations against Wang were reportedly separate: OpenAI claimed she repeatedly accessed an executive’s email. Her explanation is that she had requested the cancellation of this access nearly a year prior. She admitted to accidentally opening a sensitive email the previous month, but stated she did not forward or share its contents and reported the mistake immediately. OpenAI indicated that the access itself was not the issue.
Korbak believes the negative repercussions of this incident extend far beyond the three individuals involved. He noted that colleagues are now hesitant to communicate with external parties and reluctant to raise safety concerns internally. The three researchers have since sent a letter to the OpenAI board, urging enhanced collaboration with external safety organizations.
Separately, a recently released audit report has also raised questions about OpenAI’s safety procedures. Auditors from METR and Redwood Research identified numerous security vulnerabilities in OpenAI’s infrastructure during their review, yet OpenAI redacted nearly half of the final report’s content.
OpenAI has also published a report on a new security architecture designed to operate offline. The company stated that the identified vulnerabilities have been fixed, model training resumed in early October, and that new external collaborations will be announced soon.
