⚪ OpenAI has terminated three senior safety researchers: Jasmine Wang, Tomek Korbak, and Mikita Balesni. The stated reason is unauthorized disclosure: the researchers shared confidential infrastructure architecture data with an undisclosed third-party AI safety organization without company approval. All three had previously raised concerns internally about the pace of AI development. The Federal Trade Commission has since launched investigations into OpenAI and other frontier AI companies.
The timing matters. Safety research attrition at frontier labs is not new, but this termination follows a period of visible internal tension at OpenAI over risk evaluation timelines and deployment decisions. The FTC investigations, which are broader than this incident, add institutional pressure to a company navigating both rapid commercial growth and criticism over how much weight it assigns to internal safety voices.
The governance gap this exposes
There is no established whistleblower protection framework for AI safety researchers at frontier labs. If a researcher believes an AI system poses unacceptable risk and internal channels are not working, the options are limited: stay quiet, resign, or disclose externally, which creates legal exposure. The researchers who were terminated apparently chose disclosure to another safety organization, not to a regulator or the press. That distinction matters legally, but it illustrates the structural problem: the people best positioned to evaluate AI risk have no protected path for escalating concerns beyond their own organization.
Why this is a governance problem, not just an HR incident
Frontier AI labs are making decisions with broad societal implications. The conventional argument for self-regulation is that internal safety teams can act as a check on development pace. That argument depends on those teams having real influence and protected standing. When safety researchers are terminated for sharing concerns with a third-party safety organization, the implied message is that the internal check is not independent. That has implications not just for OpenAI but for the credibility of the self-regulation model the entire industry relies on.
What policymakers and organizations should watch
For policymakers: the EU AI Act and emerging US AI governance frameworks should consider explicit protections for safety researchers at frontier labs, similar to provisions for whistleblowers in financial services and nuclear energy. For organizations procuring frontier AI: the internal governance structure of your AI vendors is now a third-party risk factor. Assess whether the safety function at your AI providers has structural independence, including protected escalation paths.
Gigia Tsiklauri is a Security Architect and founder of Infosec.ge. Get in touch if your organization is evaluating AI vendor governance as part of its third-party risk program.