White House AI safety framework exempts open-weight models and will be kept from the public

The Trump administration detailed a new voluntary framework for vetting the cybersecurity risks of advanced artificial intelligence on Tuesday, briefing staffers from six of the world's most powerful tech companies at the White House. The framework mandates a 30-day pre-release government review of so-called closed-source or frontier AI models developed by U.S. firms, while explicitly exempting open-weight or open-source models. The White House intends to keep the framework and its underlying testing benchmarks classified.

Scope and exclusions

Multiple people familiar with the discussions told the New York Times, Politico, and Bloomberg that the review process targets closed models from developers such as OpenAI and Anthropic. These models do not publish their underlying code. The framework carves out a broad exemption for open-weight AI systems, a category that includes models from American firms like Meta, Nvidia, and Google, as well as Chinese companies such as Alibaba, DeepSeek, and Moonshot AI. A Politico source said the vetting policy would not apply to open models despite a recent surge of activity and interest in them.

The 30-day review process

The voluntary framework allows AI developers to submit new models to the federal government up to 30 days ahead of public release, according to Wired and Axios. The government will then assess their cyber capabilities according to a classified benchmarking system and share the models with federal agencies and trusted corporate partners. During the review, employee access to the models would be limited, the systems would be stored in high-security environments, and detailed access logs would be kept. Companies were reportedly encouraged to submit models as close to public release as possible rather than early-stage versions.

Secrecy and industry access

The White House does not plan to publicly release the framework. Representatives from Anthropic, OpenAI, Microsoft, Meta, Google, and Nvidia attended the Tuesday staff-level meetings, but companies that were not invited remain in the dark about its contents. A White House official told Wired that the framework is intentionally narrow and focused on the cybersecurity capabilities of the most advanced models on the market, such as Anthropic's Fable and OpenAI's ChatGPT 5.6. A second official cited national security concerns for keeping the review criteria confidential.

Criticism and the open-source gap

Safety advocates and smaller startups warned that the secretive process gives an advantage to the largest frontier labs. A person familiar with the White House discussions told Wired the framework creates an entrenchment program for big AI model providers and leaves out smaller competitors. Brad Carson, president of the nonprofit Ameri, was quoted in Wired criticizing the secrecy:

This is far too important an issue to be hidden behind a cloak of secrecy.

The decision to exclude open-source models also delays regulatory scrutiny of increasingly powerful software from Chinese developers. The New York Times reported that the rules governing the reviews are nearly complete, though what qualifies as a state-of-the-art capability or a true national security risk remains undefined. The framework is entirely voluntary, and there is currently no mechanism to compel compliance if companies stop submitting their models for evaluation.