OpenAI terminated three safety researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, following an internal investigation into alleged policy violations. The company claimed the employees mishandled sensitive information, though the researchers state they were never provided with written details regarding the specific data they supposedly leaked.
The investigation began after OpenAI security teams suspected the researchers of sharing internal model architecture details. Investigators examined personal devices and identified the use of Signal, a messaging app with disappearing messages. OpenAI stated that the use of unauthorized communication tools was not the primary reason for the dismissals.
The conflict involves access to 'chain of thought' reasoning traces, which researchers use to monitor how AI models arrive at decisions. In 2025, the three researchers co-authored a paper advocating for continued external access to these processes. The tension escalated after reports suggested that newer models were revealing less information about their internal reasoning.
Jasmine Wang faced additional accusations regarding unauthorized access to an executive's inbox. She claims she requested the removal of this access previously and reported an accidental email opening immediately. Meanwhile, an independent audit by METR and Redwood Research identified security gaps in OpenAI's infrastructure, though the company redacted nearly half of that report.
OpenAI recently implemented a new security architecture that functions without internet access and stated that training resumed in early October. The company plans to announce new external partnerships in the near future. The fired researchers have since written to the OpenAI board to advocate for increased transparency and cooperation with external safety organizations.
