White House finishes frontier AI review plan, but keeps the trigger secret

The U.S. now has a framework for bringing the most capable AI models into a pre-release government review process, but developers still do not know where the tripwire sits. Axios reported that the White House met the August 1 deadline set by a June 2 executive order, while declining to disclose the framework’s contents, audience, or rollout timing.

The framework is voluntary, but it changes the operating environment for frontier labs. It is meant to help developers determine whether models under development fall under a government engagement process. When the government gets access to a model before release, the framework is expected to define confidentiality, cybersecurity, insider-risk, intellectual-property, use, and nondisclosure requirements.

The concrete number matters: Axios says the process contemplates government access to models for up to 30 days before release. That is not a full licensing regime, but it is enough to affect launch sequencing, customer commitments, and safety review calendars for labs building models near the frontier.

The least visible piece may be the most important one. The executive order makes the benchmarking process for advanced cyber capabilities classified. The threshold for deciding which models are covered is also classified and shared with AI developers and researchers only as appropriate. That limits public accountability, but it also tells labs that cyber capability is becoming a formal gate in U.S. model governance.

OpenAI, Anthropic, and Google provided feedback on a draft before the deadline, according to Axios, and a staff-level company meeting was scheduled for Tuesday to review next steps. The question now is whether the government can give developers enough predictability without publishing a recipe for evading the tests. If it cannot, “voluntary” review may still become a practical launch constraint for the next wave of frontier systems.