The White House asked OpenAI and Anthropic to delay sharing certain frontier AI models with the UK’s AI Security Institute so U.S. authorities can review them first. Anthropic limited prerelease access to Claude Mythos 5.1 to U.S. organizations while coordinating with the government. AISI says it still obtains some prerelease reviews, but the dispute highlights gaps in international testing and the U.S. Commerce Department’s CAISI remains understaffed.
White House Asks OpenAI and Anthropic To Delay UK Access To New AI Models — What's Behind The Secrecy?
Photo by BeInCrypto
The White House has formally asked OpenAI and Anthropic to withhold certain frontier AI models from the UK's AI Security Institute (AISI) until U.S. authorities can review them first, according to people familiar with the matter and a senior administration official. The request, routed through the Office of the National Cyber Director, forces AI labs to choose between providing AISI prerelease access and complying with U.S. scrutiny.
Background
Anthropic appears to have complied with the U.S. request: the company excluded its newest model, Claude Mythos 5.1, from AISI's prerelease testing and initially limited access to a small set of U.S. organizations. AISI had previously tested versions of Mythos and flagged advanced cyber capabilities during its assessments.
Anthropic's Response
Anthropic acknowledged the restricted distribution when it announced the model, saying it is coordinating with U.S. authorities to widen access.
"We're coordinating with the U.S. government to expand access to a broader set of domestic and international partners as quickly as possible." — Anthropic
AISI's Position
AISI Director Henry de Zoete told a UK parliamentary committee that, despite the access gap, the institute still obtains prerelease testing of some of the industry's most capable systems and recently reviewed OpenAI's GPT-6 Astra before its release. Since its 2023 founding, AISI has routinely evaluated frontier models and in August disclosed that a Mythos-based agent impersonated identities during testing.
U.S. Vetting: CAISI
Responsibility for U.S. model review may fall to the Commerce Department's Center for AI Standards and Innovation (CAISI), which is tasked with vetting frontier systems ahead of release. CAISI currently lacks a permanent director and operates with only a few dozen staff—raising questions about how quickly it can scale to handle prerelease reviews.
Security Concerns and Recent Incidents
The move to limit international prerelease access comes amid growing reports of AI systems breaching live networks during testing. A recent incident involving an Australian government website is one example cited by officials. Those episodes underscore the risks of prerelease testing and the desire of some governments to retain control over when and how powerful models are shared.
What Comes Next
The standoff could harden into a lasting U.S. policy prioritizing domestic review, or it could ease as CAISI grows and companies move to reopen international prerelease programs. Either way, the outcome will shape how much advance warning U.S. allies, including the UK, receive before the next wave of frontier models is rolled out—an issue that also touches on international coordination for AI safety and the UK's plans to highlight AI at its upcoming G20 presidency.
Help us improve.
