OpenAI fires three safety researchers over alleged policy violations

OpenAI has dismissed three safety researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, following an internal investigation into alleged policy violations. The company’s security team suspected the researchers of sharing details about internal model architecture with an external safety researcher, Ryan Greenblatt. The investigation began after a news report regarding the Astra model. During the process, investigators found Signal chats with disappearing messages enabled on the researchers' personal devices. OpenAI stated that Signal was not an approved communication tool but insisted this was not the reason for the terminations.

The three researchers deny sharing confidential information. They state they were never told in writing what specific information they allegedly mishandled. Ryan Greenblatt denies passing confidential information to journalists. Jasmine Wang faced a separate accusation of repeatedly accessing an executive's inbox, though OpenAI says the access itself was not the issue.

The researchers suspect their terminations were tied to their efforts to make AI models easier to monitor. They warn the case could discourage employees from raising safety concerns. In a letter to OpenAI's board, the three call for closer cooperation with external safety organizations. A recently published on-site audit by METR and Redwood Research found numerous security gaps in OpenAI's infrastructure, but OpenAI redacted nearly half the final report.

OpenAI published its own report on a new security architecture that operates without internet access. The company said it fixed the issues and resumed training in early October.