White House to Keep AI Model Safety Testing Scheme Classified

The White House does not plan to publicly reveal a new scheme for evaluating advanced artificial intelligence models. The voluntary verification mechanism has global implications for AI safety, but details will be disclosed only to companies involved in the process. Ahead of this, the administration held a meeting with AI industry representatives to discuss the completion of the framework. Companies not invited to the meeting remained unaware of its content. The framework was developed amid industry struggles with cyberattacks and rapid progress by Chinese AI developers, reigniting debates on open models. It is unclear how trusted partners will get early access to advanced models and whether friendly foreign governments will be included. Unconfirmed reports say open models were discussed and Nvidia staff participated; last week Nvidia CEO Jensen Huang visited Washington, meeting Commerce Secretary Howard Lutnick and President Donald Trump. The voluntary scheme, outlined in a June executive order, describes a mechanism for AI developers to interact with authorities and sets criteria for which models fall under the document. US government bodies get access to advanced models 30 days before release, testing for privacy, cybersecurity, insider risks, intellectual property issues, and non-disclosure requirements. The procedure for assessing advanced AI cybersecurity capabilities is classified, as stated directly in the document.