The White House has asked OpenAI and Anthropic to refrain from sharing their new artificial intelligence models with the British government's testing agency. This information was reported by Politico, citing sources.
According to the publication, the request came from the U.S. National Cyber Director's Office. The administration aims to first evaluate the models and ensure the safety of American systems before sharing them with partners.
Politico notes that this situation presents a dilemma for the companies involved. They must either withhold their models from the UK’s AISI, which previously had early access, or risk a conflict with the White House.
Journalists claim that Anthropic has agreed to the request, withholding the Claude Mythos 5.1 model from AISI. Access to this model has only been granted to select organizations in the United States.
Earlier in September, AISI Director Henry de Zute informed a parliamentary committee that the institute had not received a new model for evaluation. He did mention, however, that AISI had tested OpenAI's GPT-6 Astra prior to its release.
“The strong partnerships maintained by AISI allow us to test a range of models that demonstrate significant advancements across the industry,” de Zute stated.
Established in 2023, AISI regularly tests new models from leading AI labs on a voluntary basis before they are publicly launched.
Politico linked the White House's request to ongoing discussions about how the U.S. should manage powerful AI models following incidents during testing phases. The publication cited examples such as breaches of external resources, including the Australian government’s website.
It is worth noting that on September 21, UN experts urged proactive measures to prepare for potential loss of control over AI.