White House Presses OpenAI and Anthropic to Give U.S. Reviewers First Look at New Models

The White House has asked OpenAI and Anthropic to route new frontier models through U.S. government review before handing them to the U.K.'s AI Safety Institute, according to reporting from The Decoder. The arrangement would put Washington's agencies at the front of the line, ahead of British testers who have worked with both labs since the institute opened in 2023. It signals that the administration intends to treat access to unreleased models as a matter of U.S. oversight rather than routine international scientific cooperation.

The U.K. AI Safety Institute, founded after the 2023 Bletchley Park summit, has run pre-deployment evaluations on models from both companies under voluntary agreements. Those tests cover dangerous capabilities such as cyberoffense, bioweapon uplift and autonomous replication, and the institute publishes summaries of some findings. Putting a U.S. review gate in front of that pipeline would delay or reshape a channel that has become one of the few formal ways governments inspect models before release.

Neither OpenAI nor Anthropic has publicly confirmed the request, and the White House has not issued a written policy describing it. Both companies have existing ties to U.S. agencies: OpenAI and Anthropic signed agreements with the U.S. AI Safety Institute, part of the National Institute of Standards and Technology, and Anthropic has worked with the Defense Department and intelligence community on model deployment. The new instruction appears to extend that relationship into a de facto clearance step for foreign testers.

The timing matters because the U.S. and U.K. have spent two years building parallel testing regimes and promising interoperability. In 2024 the two countries signed a memorandum of understanding to align their safety institutes, including shared evaluation methods and joint research. A first-look demand from Washington cuts against that stated goal, leaving British officials to either accept second position or negotiate parallel access.

For the labs, the request creates a scheduling problem. Model releases are already gated by internal red-teaming, and adding a U.S. government review before U.K. testing lengthens the window between training completion and public launch. It also raises the question of what happens when American and British evaluators reach different conclusions about a model's risk level, and who gets to decide whether a release proceeds.

What to watch next is whether the arrangement is formalized in writing and whether the U.K. AI Safety Institute confirms any change to its access. If the institute loses early access to U.S. frontier models, other allied testers in the EU and Japan could face similar limits, pushing more of the world's pre-deployment scrutiny behind Washington's review process.