OpenAI Dismisses Three Security Researchers Over Alleged Leaks

Three researchers from OpenAI have been dismissed for alleged leaks to an external organization, and a fourth is said to have left the company. Key roles within the security and alignment teams are involved, as several points remain unconfirmed and the individuals had recently raised concerns about the risks of AI.

No established link and unconfirmed identities

The Wall Street Journal does not establish any connection between the work done with METR and Redwood Research and the dismissals. OpenAI has not confirmed the names of the individuals involved, and it is not specified what information may have been transmitted or to which organization. Additionally, an anonymous account on X claims that a fourth security researcher, David Robinson, left OpenAI shortly thereafter.

OpenAI ends three collaborations for alleged violations

OpenAI has terminated the collaboration of three researchers for alleged leaks of confidential information to an external AI security organization, according to the Wall Street Journal. A spokesperson for the company stated that an internal investigation confirmed violations of rules regarding the handling of sensitive information. The individuals identified by the journal are Jasmine Wang, Tomek Korbak, and Mikita Balesni. Korbak was part of the security team, while Wang and Balesni worked on alignment.

External points of contact and public alerts on risks

Tomek Korbak served as the technical liaison between OpenAI and METR as well as Redwood Research, two organizations tasked with analyzing how OpenAI's AI agents manage to bypass security measures and access external platforms such as Hugging Face. In September, all four researchers publicly expressed their concerns about the dangers associated with AI. Korbak indicated he was dissatisfied with several decisions made by OpenAI. Mikita Balesni believes that the chance of AI eliminating all of humanity exceeds ten percent. Jasmine Wang signed a petition calling for a slowdown in AI advancement. David Robinson admitted that the competition to create self-improving AI could be seen as absurd.