White House finalizes frontier-AI review framework, exempts open-weight models from vetting

The White House has finalized a voluntary security-review framework for frontier AI models that exempts open-weight systems from government vetting, according to people familiar with the matter. The decision concentrates scrutiny on closed models from the largest US labs. The administration conveyed it to OpenAI, Anthropic, Google and Meta on Tuesday.

The exemption is a consequential choice. It leaves openly released model weights, which anyone can download, modify and run, outside a review meant to judge whether a system could help discover software vulnerabilities or carry out cyberattacks.

Under the program, participating developers can give the federal government up to 30 days of early access to a new frontier model before it reaches other partners, so agencies can assess its offensive-cyber potential. The framework was finalized under a June 2, 2026 executive order.

It carries no enforcement mechanism. The program is voluntary, not a mandatory licensing or preclearance regime, and companies face no penalty for declining to take part.

The administration does not plan to publish the framework itself, and the specific benchmark and qualifying threshold that decide which models fall in scope are expected to remain classified. That secrecy makes it hard for outsiders to judge how demanding the bar is, or how many models it will actually cover.

The approach reflects a bet that early access plus lab cooperation can flag dangerous capabilities without the friction of formal regulation. Whether it holds depends on labs volunteering their most capable systems and on a classified threshold no one outside government can see.