According to a report by the American 'Politico News Network' on September 24th, a source familiar with the matter and a senior US government official revealed that the White House has asked OpenAI and Anthropic to delay providing new models to British testing agencies until the US completes its review process.
The report indicates that this requirement proposed by the U.S. National Cyber Director’s Office puts companies in a dilemma: they can either refuse to provide their models to the British Artificial Intelligence Safety Institute—an institution with one of the world’s most abundant resources for government testing, which previously had the privilege of receiving models first; or they risk provoking dissatisfaction from the Trump administration.
The aforementioned senior U.S. government official stated that the White House hopes for the U.S. side to first review the model and take measures to ensure the security of the U.S. system, before sharing the model with U.S. partners. “Because they are U.S. companies, we have always adopted this policy regarding every new, cutting-edge model that is released.”
The White House did not immediately respond to the request for comment.
According to reports, Anthropic has apparently agreed to this request. The company did not provide the Claude Mythos 5.1 model to the UK Artificial Intelligence Safety Institute. In their announcement, they stated that the model is “available only to a limited number of US institutions”.
Anthropic wrote in their blog post about releasing this model: “We are coordinating with the U.S. government to extend access to more domestic and international partners as soon as possible.”
In early September, Henry de Zoete, director of the UK Institute for Artificial Intelligence and Security, confirmed in a letter to the UK parliamentary committee that it was not possible to access Anthropic’s models.
But he said, “We maintain close contact with all leading AI research companies and continue to have access to some of the world’s most advanced models in advance.”
He pointed out that the agency completed tests before the official release of OpenAI’s GPT-6 Astra model. “The reliable partnership with the UK Artificial Intelligence Safety Institute allowed us to test a series of models that represent a significant leap in industry capabilities.”
OpenAI's GPT-6 Astra Artificial Intelligence Large Model IC Photo
OpenAI has not responded to requests for comments from "officials."
The company said this week that the United States Artificial Intelligence Standards and Innovation Center (CAISI) should lead the development of international standards for monitoring AI developments and safety, and cooperate with national AI safety agencies such as the UK Artificial Intelligence Safety Institute.
According to reports, in the United States, Congress and the White House are working to allocate more resources to CAISI, an organization under the Department of Commerce. This institution is considered the core entity for AI management in the US. However, it currently has no fixed leader. CAISI has only a handful of technical staff, yet it is tasked with evaluating cutting-edge models that continue to emerge.
While the Trump administration is making these demands, the White House is trying to determine how the government should respond to powerful new AI models that invade external entities during testing. For example, there was a recent incident where an AI agent developed by OpenAI invaded the Australian government’s website.
