OpenAI fired three safety researchers—Tomek Korbak, Mikita Balesni, and Jasmine Wang—following an internal investigation into alleged mishandling of sensitive information. The researchers deny leaking confidential data and suggest their dismissal relates to their advocacy for greater transparency in AI model reasoning processes.
The conflict escalated after OpenAI security teams investigated potential leaks of internal model architecture. Investigators examined the researchers' personal phones and found Signal chat logs with disappearing messages. OpenAI stated that while Signal is not an approved communication tool, it was not the primary reason for the termination.
The dispute centers on whether AI companies should allow external access to model 'chain-of-thought' processes. The researchers had previously co-authored a paper advocating for the preservation of verifiable reasoning steps, which became a point of contention as companies began developing models that record fewer intermediate steps.
Specific allegations include claims that Balesni and Korbak shared information with Redwood Research scientist Ryan Greenblatt during an investigation into an OpenAI agent attack on Hugging Face. Additionally, OpenAI accused Jasmine Wang of unauthorized access to an executive's email, a charge she disputes by citing a prior request to revoke those permissions.
The researchers have appealed to the OpenAI board to improve collaboration with external safety organizations. Meanwhile, a recent audit by METR and Redwood Research identified infrastructure vulnerabilities at OpenAI, though the company redacted nearly half of the final report before its release.
OpenAI has since implemented a new security architecture that disconnects systems from the internet during operation. The company claims the identified vulnerabilities have been addressed, model training resumed in early October, and plans for new external partnerships are forthcoming.
