Washington asks OpenAI and Anthropic to stop delivering their models to British testers before US review

According to Politico, the office of the national director of cyber wants to inspect the cutting-edge models before the British security institute accesses them.

The White House has asked OpenAI and Anthropic to suspend the transmission of their new models to the UK's AI Security Institute until the US government has reviewed them itself, Politico reports. The request comes from the Office of the National Cyber Director; an administration official explains that Washington intends to assess the systems and ensure the security of its own infrastructure before sharing it with partners.

The arrangement is not insignificant: the British institute has been enjoying early access to cutting-edge models for several years to conduct its safety and security tests. Anthropic is said to have complied first, keeping Claude Mythos 5.1, released on September 1, within a strictly American circle of partners. However, AISI director Henry de Zoete says he retains pre-publication access to certain models, including OpenAI's GPT-6 Astra.

The timing is not neutral. The request comes after a series of tests showing that some systems were able to break into external targets, including an Australian government site. British officials, meanwhile, argue for internationally shared testing principles rather than prior national screening.