The Trump administration has finalized its framework for assessing the cybersecurity risks of advanced AI models. The plan reportedly covers only the top tier of proprietary models, skips open-weight systems entirely, and leaves its central concept, a national security risk, undefined.
The administration shared details of the framework with OpenAI, Anthropic, and other leading AI labs during a meeting at the White House on Tuesday. The program is voluntary: developers would submit their most capable proprietary models for review rather than being compelled to do so, according to reporting from Politico, Axios, The Verge, and others. The White House has not publicly released the criteria, the list of covered models, or the process by which reviews would happen.
The exclusion of open models is the most consequential detail to emerge. The plan reportedly has no interest in testing open-weight systems, the class of models that anyone can download, inspect, and modify. The New York Times reports the administration does not plan to review such models, although that position could shift as the technology evolves. The decision mirrors an ongoing debate inside the industry: some researchers argue that open weights now approach frontier capability while lacking the safety layers that labs build into their flagship products.
Equally notable is what the document does not contain. According to The Verge, the framework never defines what constitutes a national security risk, leaving unclear what threat level would trigger a review, what findings would be reported, and what, if anything, would happen to a model that failed. Without that definition, a voluntary review program has no meaningful threshold, which makes it hard to assess whether it is a genuine security mechanism or a public-relations gesture.
The secrecy compounds the problem. The public learned about the framework through leaks and lab briefings, not through publication. The administration says details will follow, but for now the only people with a complete picture are the companies being asked to participate.
The framework lands at a moment when the open-weight question has moved from academic to urgent. Governments are watching frontier and open models converge in capability, and the gap between what such systems can do and how they are governed has become the central policy tension in AI. A review program that exempts open models and declines to define its own terms does little to close that gap.
Sources: Trump’s AI testing plan is limited and vague (The Verge, Aug 5, 2026); White House AI vetting plan to exempt lower-cost ‘open’ models (Politico, Aug 4, 2026); Trump White House Readies AI Framework to Review Security Risks (The New York Times, Aug 4, 2026); The White House Is Keeping Its AI Cybersecurity Framework Secret (Wired, Aug 4, 2026)