OpenAI Fires Three Safety Researchers Amid Transparency Dispute

On October 25, three OpenAI safety researchers fired on suspicion of leaking confidential information shared their side of the story. Tomek Korbak, Mikita Balesni, and Jasmine Wang denied leaking sensitive data, challenged the official justification for their dismissal, and speculated that the incident was related to their longstanding advocacy for greater transparency in AI systems.

Following an internal investigation, OpenAI determined that the three individuals violated its sensitive information management policies and terminated their employment. Prior to being fired, all three had publicly voiced opinions regarding AI risks.

IT Home noted that in interviews with The Wall Street Journal, the researchers rebutted the official narrative. OpenAI told the newspaper that the dismissals were unrelated to the safety concerns they had raised. However, the researchers stated that the company never provided written notice specifying which information had been mishandled. They also expressed concern that this incident might deter independent security auditors from collaborating with AI companies.

The Wall Street Journal’s report reconstructed the origins of the conflict. The OpenAI security team suspected that the three had leaked details about internal model architectures to an external security researcher and demanded that they unlock their personal phones. Investigators discovered Signal chat logs with disappearing messages enabled on the devices. A week later, all three were simultaneously fired. OpenAI stated that Signal is not an approved communication tool but emphasized that this was not the reason for the termination. Korbak countered that using disappearing messages is a standard security practice in Silicon Valley.

Core Dispute: Should External Verification of AI Reasoning Steps Be Allowed?

At the heart of this dispute lies a profound technical question: should AI companies allow external parties to inspect models' thought processes? Reasoning models record intermediate steps in an internal "chain-of-thought" scratchpad. Companies and security researchers rely on these records to understand the logic behind a model's decisions.

In a 2025 paper, Balesni, Korbak, and Wang collaborated with top researchers from multiple AI labs to call for the continued preservation of this verifiability capability. This issue became more urgent this summer as several AI companies began internally discussing a new class of models—those that take longer to think but record fewer intermediate steps. Balesni raised objections within the company by forwarding a relevant proposal from Ryan Greenblatt, Chief Scientist at Redwood Research (now known as METR).

Korbak also maintained frequent communication with Greenblatt. The two collaborated via Slack and Signal to jointly analyze a high-profile incident: the attack on Hugging Face by an OpenAI agent. Greenblatt was commissioned by METR to lead this independent investigation.

An Article About Astra Triggered the Internal Investigation

On September 1, The Information published a report stating that OpenAI’s new model, Astra, was more capable but exposed less of its reasoning process. This collaboration subsequently became the catalyst for trouble. At the time, OpenAI Chief Scientist Jakub Pachocki criticized the article on X (formerly Twitter) as logically flawed, yet he himself had previously warned that the industry was racing to develop models that are difficult to monitor.

According to The Wall Street Journal, weeks later, OpenAI launched an internal leak investigation. Balesni said investigators identified Greenblatt as the recipient of the leaked information and requested access to their Signal chat history. Greenblatt denied passing any confidential information to journalists.

The allegations against Wang stood separately: OpenAI claimed she accessed an executive’s email account multiple times. Her explanation was that she had applied to revoke this permission nearly a year earlier; last month, she accidentally opened a sensitive email, did not forward or disseminate its contents, and immediately reported the mistake. OpenAI stated that the permission itself was not the issue.

Korbak argued that the negative impact of the incident extends far beyond the three of them. He noted that colleagues are now afraid to communicate with outsiders or raise safety concerns internally. The three researchers have written a letter to the OpenAI Board of Directors, calling for stronger cooperation with external security organizations.

A newly released on-site audit report also questioned OpenAI’s security processes. Auditors from METR and Redwood Research discovered numerous security vulnerabilities in OpenAI’s infrastructure during their review, but OpenAI redacted nearly half of the final report.

OpenAI has also published a new security architecture report describing a system that operates without internet connectivity. The company stated that the relevant vulnerabilities have been fixed, resumed model training in early October, and announced that it would soon reveal several new external partnerships.