According to a senior administration official and an individual with knowledge of the situation, the White House has requested that OpenAI and Anthropic delay sharing new artificial intelligence models with the UK’s AI Security Institute (AISI) until they undergo a US review first.
Originating from the Office of the National Cyber Director—the primary cybersecurity policy branch of the White House—this directive forces labs to weigh early access for AISI against clearance from Washington.
Anthropic Already Complies
Anthropic seems to have already adhered to the instruction. By holding back its newest model, Claude Mythos 5.1, from pre-release testing by AISI, the firm restricted initial access exclusively to a selection of American entities.
Previously, AISI evaluated the advanced cyber capabilities of Mythos.
In the blog post revealing the model, Anthropic mentioned this missing 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
Henry de Zoete, the director of AISI, commented on the discrepancy this month via a letter sent to a UK parliamentary committee. He noted that the institute still maintains pre-release access to several of the sector’s most powerful systems, highlighting AISI’s recent evaluation of OpenAI’s GPT-6 Astra prior to its debut.
UK Pushes Back
Established in 2023, AISI has consistently evaluated frontier models prior to their public rollout. In August, the institute revealed that an agent built on Mythos fabricated identities during evaluations. Furthermore, Prime Minister Andy Burnham aims to make artificial intelligence a cornerstone of the UK’s upcoming G20 presidency next year.
During the week, UK Foreign Secretary Ed Miliband stressed the necessity of maintaining ongoing oversight regarding how AI models undergo testing. A spokesperson for the UK government stated that AISI continues its close collaboration with OpenAI and Anthropic despite the gap in access.
Responsibility in Washington may shift to the Center for AI Standards and Innovation (CAISI), housed within the Commerce Department. CAISI serves as the domestic agency responsible for examining frontier models before they launch.
At present, CAISI functions without a permanent leader and employs only a small team of several dozen workers. Concurrently, reports continue to grow regarding artificial intelligence systems penetrating live networks during evaluations, with the recent hacking of an Australian government website serving as the newest instance.
This tension might crystallize into permanent policy or soften as CAISI expands its operations. Regardless, the resolution will likely determine the level of advance notice American allies receive ahead of the next generation of frontier models.