US Plans AI Security Review for 'Closed' Models, Exempts 'Open' Ones

The Trump administration is finalizing a security review process that would apply to 'closed' frontier AI models before their public release, while explicitly exempting 'open' models from the same scrutiny. Under the framework, developers of closed, state-of-the-art models flagged for national security risk would face a roughly 30-day government pre-release review, during which employee access would be restricted and the model would be stored in a high-security environment with detailed access logs.

The plan followed a White House meeting on Tuesday that included OpenAI, Anthropic, Google, Nvidia, Microsoft, and Meta, the companies most likely to be covered by the closed-model review given they build the era's most capable proprietary systems. Open-weight models are carved out entirely: the framework states nothing in it should be read as restricting open models once released, a distinction officials frame as necessary to keep the US competitive with China's fast-growing open-source AI push.

The approach effectively creates two regulatory tracks, tighter oversight for the handful of companies building the most capable closed systems, and a lighter touch for open-source developers, a split that could influence where future frontier model development happens. Details of the full framework have so far been shared only with the companies that attended the meeting, not published publicly, leaving open questions about which models meet the 'national security risk' threshold and how review timelines might affect product launch schedules.