White House Weighs Pre-Release Vetting for AI Models

The White House is exploring a formalized process that would give federal agencies early access to powerful AI models for security testing before companies release them publicly. Officials from the Office of the National Cyber Director and the White House Office of Science and Technology Policy have briefed major labs as part of those discussions.

Those briefings, described by technology outlets and people familiar with the matter, covered proposals ranging from voluntary early-sharing agreements to a more structured executive-level review pathway. The conversations aim to map how agencies would assess cyber risks and other national-security concerns without immediately blocking commercial releases.

The talks gained urgency after recent model rollouts prompted concern inside government. The administration asked OpenAI to stagger the launch of its newest model and limit initial access to a small set of approved partners while agencies perform security assessments. Companies and officials say the approach is meant to buy time to test capabilities and controls.

Anthropic’s recent decision to restrict access to its frontier models and a rare Commerce Department directive that followed the Mythos release have also pushed officials to consider a formal review mechanism. Those incidents underscored how fast capabilities are evolving and how models can raise cyber and infrastructure risks.

President Trump signed an executive order in early June that lays out a federal framework to strengthen AI cybersecurity and innovation, and it assigns timelines for agencies to develop testing and oversight procedures. The order instructs multiple agencies to help determine whether specific systems qualify as “covered frontier models.”

At the same time, the White House fact sheet for the order states explicitly that the directive “shall not be construed to authorize” creation of mandatory government licensing, pre-clearance, or permitting requirements for AI development and release. That tension — between asserting federal oversight and avoiding formal pre-market bans — frames much of the policy debate.

People familiar with the talks say one practical model under consideration would give vetted agencies early, controlled access to model snapshots or test APIs so cybersecurity teams from CISA, the Pentagon, and other bodies could run red-team exercises and code analyses. Some government and industry officials have discussed approving access “customer by customer” during preview periods.

Industry response has been mixed but largely cooperative so far. Major labs have agreed to some voluntary limits on distribution and to briefings with federal officials, while also warning that heavy-handed rules could slow innovation and push risky work overseas. The White House has also been urging other firms to voluntarily submit their most advanced systems for review.

Legal and technical experts say implementing a pre-release pathway raises thorny questions about classified testing, proprietary code, and how to weigh national-security risks against economic competitiveness. Advocacy groups and some lawmakers want independent audits and transparency requirements to be part of any review.

Policymakers are also watching international models. Some officials have cited the U.K.’s approach to assessing frontier systems as a reference point as Washington debates whether government-led evaluations should be routine or reserved for the rarest, highest-risk releases. Those comparisons are shaping proposals inside the White House.

For now, the administration’s next steps are uncertain. Officials continue talks with the major labs and are weighing how to convert briefings and voluntary cooperation into an operational framework that meets the deadlines and obligations established in the June executive order. Any formal announcement, White House aides say, would come from the administration once the details are finalized.