Three Fired OpenAI Safety Researchers Deny Leaking, Question Dismissal
Three OpenAI safety researchers fired for alleged leaks—Tomek Korbak, Mikita Balesni, and Jasmine Wang—are denying that they disclosed confidential information and questioning the official reason for their dismissal. They suggest the action may be linked to their public advocacy for greater transparency in AI systems. OpenAI said an internal investigation found that the three violated its sensitive-information policies, but the researchers say they were never told in writing which specific information was mishandled. The trio has written to OpenAI's board calling for stronger collaboration with external safety researchers.
According to The Wall Street Journal, OpenAI's safety team suspected the researchers of leaking details of internal model architecture to an outside safety researcher and asked them to unlock their personal phones. Investigators found Signal chat histories with disappearing messages enabled. The three were fired a week later. OpenAI said Signal is not an approved communication tool but emphasized that this was not the reason for the dismissals. Korbak countered that disappearing messages are a routine security practice in Silicon Valley.
At the heart of the dispute is whether AI companies should allow outsiders to inspect models' reasoning processes. Reasoning models store intermediate problem-solving steps in an internal "chain-of-thought" buffer, which companies and safety researchers use to understand how models reach conclusions. In 2025, the three researchers co-authored a paper with investigators from multiple AI labs calling for preserving this auditable capability. This summer, as AI companies began discussing models that think longer but document fewer intermediate steps, Balesni raised objections internally. Korbak had also been collaborating with Redwood Research chief scientist Ryan Greenblatt on analysis of an incident involving an OpenAI agent attacking Hugging Face. A September 1 report by The Information, saying OpenAI's Astra model exposes less reasoning, is described as the flashpoint.
Investigators identified Greenblatt as the alleged recipient of leaked information, but he denied passing confidential material to reporters. On the separate allegation against Wang, OpenAI said she repeatedly accessed an executive's mailbox; she explained that she had requested removal of that access nearly a year earlier and had accidentally opened one sensitive email in September, reporting the mistake immediately without forwarding it. Korbak said the affair has made colleagues reluctant to communicate with outsiders. A separate field audit by METR and Redwood Research questioned OpenAI's security processes; OpenAI published a new security architecture report, saying vulnerabilities have been fixed and model training resumed in early October.
