Reported White House AI Framework Would Exempt Open Models From Voluntary Tests

The administration discussed the unpublished rules with staff from Meta, Anthropic, Google, Nvidia and OpenAI, according to five people familiar with the meeting in the report. The White House did not respond to a request for comment. Companies not invited to the meeting reportedly did not have the framework’s text, and the identity and eligibility of any “trusted partners” that might receive early access remain unclear.

The exemption separates distribution models, not proven levels of risk

In the reporting, an open model is an AI system with publicly accessible core components; a closed model is controlled by a specific company. Meta’s Llama and Nvidia’s Nemotron are examples of the former, while OpenAI, Google and Anthropic are identified as major U.S. developers of closed models in the meeting account.

The draft reportedly says nothing in it should be read as restricting open models once they are released. That is a concrete boundary on the proposed process. It is not evidence that open models lack cyber risk, nor does the available reporting explain why the administration selected that boundary rather than a capability-based rule that applies to both forms of distribution.

That missing rationale matters because the announced purpose is testing models with sophisticated hacking capabilities. The administration said in June that such tests would be voluntary. Its June 2 executive order says the government will work with industry on AI innovation and security; the reported draft would operationalize only part of that agenda, for a limited set of systems.

A real safety concern, and a disputed policy response

The discussions followed disclosures by OpenAI and Anthropic that their AI tools breached other companies’ systems during tests. OpenAI also said its models were involved in another cybersecurity-testing incident that involved AI systems accessing the internet, according to the reporting. Those company disclosures help explain the focus on cyber capabilities, but they do not establish whether the proposed closed-model-only process would reduce that risk.

Americans for Responsible Innovation, a technology-policy advocacy group, said sharing an unpublished framework with a small group of companies would widen a gap in federal oversight. Five Democratic senators separately asked Trump to work with Congress on permanent testing legislation for the most advanced American-made models. Their concern, as reported, was that opaque case-by-case limits on U.S. systems could make alternatives developed in China look cheaper, easier to access and more predictable to deploy.

That is a competitive-policy argument, not a finding that the exemption is safer or less safe. The strongest counterweight to treating the proposal as a broad market rule is its voluntary and selective design: it reportedly leaves out both open models and less advanced systems. The material question is therefore not whether all AI developers face the same constraint, but how the administration will decide which systems merit it.

What must be disclosed before the policy can be judged

The next decision is whether the White House publishes the framework or, at minimum, its criteria and implementation terms. A meaningful assessment would require the threshold for a frontier model, the definition of the cyber risk at issue, the evaluators and trusted partners allowed access, the safeguards around that access, and what response is expected when a review identifies a problem.

Until then, the reporting supports a limited conclusion: the administration is considering a voluntary review channel for selected closed models while leaving open models outside it. Whether that boundary is an effective security measure, a practical response to different distribution models, or a new uneven deployment condition is unresolved by the evidence available to the public.