- Lead. The Trump administration is finalising a voluntary framework with OpenAI, Anthropic and Google that would give federal agencies up to 30 days to evaluate new frontier AI models for national security implications before public release, with an announcement expected before August 1.
- Fact. The deal flows from a June 2 executive order directing Treasury, Defense, and Homeland Security to establish an AI benchmarking process within 60 days; the evaluation benchmarks themselves are classified.
- Stake. Meta is excluded from the agreement, leaving a major developer of capable, open-weight models operating outside the framework and creating a structural gap in any de facto oversight system.
What the Framework Does
Under the agreement taking shape, developers of covered frontier models would voluntarily provide US government agencies with early access for up to 30 days before releasing a model to other partners or the public. The period is framed as a maximum access window rather than a mandatory waiting period—agencies cannot legally block a release—but the arrangement gives national security reviewers time to flag risks, recommend changes, or escalate concerns before a model reaches the open market. The National Security Agency would classify which AI systems qualify as “covered frontier models” through a benchmarking process the White House has kept secret. An announcement is expected before August 1. The Eastern Herald reported the details of the talks.
Voluntary, but With Enforcement Levers
The “voluntary” label describes the framework’s legal structure, not its practical weight. Reporting on the talks describes informal enforcement tools available to the administration including export control threats, delayed launch approvals, and direct outreach from cabinet officials to lab leadership. AI labs are dependent on semiconductor export licences to access the high-end chips that train and run frontier models; that dependence gives the executive branch meaningful leverage even without formal mandates.
The framework parallels legislative efforts in Congress to create a formal oversight structure for the most capable AI systems—efforts that have moved slowly relative to the pace of model development. The administration’s approach sidesteps the legislative bottleneck by working directly with the companies most likely to release systems with national-security relevance.
The Meta Gap
Meta’s absence from the deal is the most structurally significant feature of the framework. The company develops and releases capable open-weight models under the Llama family, making them freely available in ways that closed-model agreements cannot easily address. Unlike OpenAI’s GPT series, Anthropic’s Claude, or Google’s Gemini, Llama models can be downloaded, modified, and deployed by anyone worldwide—including state actors and adversaries—once released. A framework covering only the three closed-model developers provides a 30-day window on some frontier systems while leaving others outside any review entirely.
The administration has indicated it may seek to expand the framework to include Meta in a later phase, but no timeline has been announced. Industry observers note that bringing Meta into a voluntary agreement would require a different enforcement logic: the company cannot delay the redistribution of a model once it is publicly posted. Whether that structural difference leads to a distinct treatment of open-weight models—or simply to their exclusion from any review regime—is the unanswered question at the centre of the framework’s design.