The White House completes a voluntary frontier-AI cyber rulebook

Subject: Leading labs are reviewing a framework for pre-release government access to high-risk models.

Preview: The proposed framework could give US officials up to 30 days to evaluate frontier cyber capabilities without creating mandatory licensing.

Governments want earlier insight into dangerous capabilities, while AI labs want to avoid a general system of pre-approval.

Meta, OpenAI, Google and Anthropic are reviewing a voluntary White House framework. It would allow up to 30 days of pre-release government access for frontier cyber evaluation; benchmarks and thresholds remain classified.

A defined review window could improve coordination before a risky launch and establish common evidence for cyber readiness. It may also influence procurement and international safety norms.

The executive approach explicitly stops short of mandatory licensing or broad preclearance. Voluntary participation, classified criteria and enforcement limits will determine how strong the framework becomes.

AI providers should prepare reproducible cyber evaluations and disclosure playbooks. Enterprise buyers should ask whether high-risk models have undergone independent or government testing.

The rulebook is an attempt to add a safety checkpoint without turning every model release into a licensing decision.