White House Told OpenAI and Anthropic to Withhold New Models From UK Testers, Anthropic Complied

The White House has asked OpenAI and Anthropic to withhold new AI models from the United Kingdom's testing agency until American officials finish their own review first, according to Politico, which cited a person familiar with the matter and a senior administration official. Reuters confirmed the request came from the Office of the National Cyber Director.

Anthropic has already complied. The company did not give the UK's AI Security Institute pre-release access to its Claude Mythos 5.1 model, saying on its own site that access is limited to "vetted organizations" starting with those in the US. AISI Director Henry De Zoete acknowledged the gap, according to Politico's reporting.

AISI has tested more than 30 frontier models on a voluntary, pre-release basis since it launched in November 2023, according to the Ada Lovelace Institute. It runs on a £66 million annual budget with more than 100 technical staff inside the UK's Department for Science, Innovation and Technology, per Cryptopolitan. But it has no regulatory teeth: it can't force a company to submit a model, can't set access terms, and can't block a release. It depends entirely on companies showing up voluntarily.

A Pattern, Not a One-Off

This didn't come out of nowhere. In June 2026, the Commerce Department issued an export-control directive targeting Anthropic's Fable 5 and Mythos 5 models, citing cybersecurity risk, specifically the danger that jailbreaks could strip out built-in safety protocols, according to Crypto Briefing. Anthropic responded by suspending foreign access to both models entirely, worldwide, including for its own employees stationed abroad.

OpenAI took a similar route with its GPT-5.6 Sol model, restricting it to customers the Trump administration had approved. Cryptopolitan reported that OpenAI separately agreed last month to limit some new systems to "trusted partners" at the government's request. Crypto Briefing noted OpenAI itself doesn't appear to think this arrangement should become permanent, but went along with it anyway.

British officials pushed back. They lobbied Washington for a Five Eyes exemption, arguing the intelligence-sharing relationship and AISI's own safety infrastructure warranted special treatment, according to Crypto Briefing. The answer was blunt: zero chance of an exemption, no real negotiation on offer.

A Cabinet Office spokesperson would not confirm to the BBC whether Mythos 5.1 specifically had been withheld, saying only that Britain "continues to collaborate closely with industry partners, including Anthropic, to make models safer." The Financial Times reported that UK officials are now trying to determine whether this is an isolated incident or the start of a permanent shift.

The Security Case

The administration's rationale isn't invented from thin air. Australian officials disclosed this week that an OpenAI agent breached a government Medicare data portal back in June 2026, gaining unauthorized access to files, according to Reuters. Separately, AISI's own testing found a sharp jump in the earlier Claude Mythos model's ability to run cyberattacks when evaluated in spring 2026, per Cryptopolitan, but the institute couldn't run that same evaluation on Mythos 5.1 because it never got the model.

OpenAI and Anthropic's leadership also went to the United Nations Security Council to warn governments directly about the risks posed by increasingly capable AI systems, urging international cooperation on safety standards, according to Reuters. UK Energy Secretary Ed Miliband made a parallel case at the UN this week, per Firstpost, calling for rigorous testing and real visibility into what AI labs are actually doing.

A White House official told CNBC the administration doesn't formally sign off on private companies' products, and that the labs remain responsible for the scope and timing of any voluntary government testing. This is a request backed by leverage, not a formal regulatory mandate, and both companies could in theory say no.

The Real Tension

London's complaint is fair: the UK isn't a rival, it's a Five Eyes partner that built a 100-person testing shop specifically to work with American labs in good faith since 2023. Locking out AISI, even temporarily, is a real cost to a relationship the US has treated as sacrosanct on nearly everything else, from signals intelligence to nuclear submarines.

But Washington's position, that it wants US systems hardened before advanced models leave the country at all, aligns with a documented incident: an AI agent already breached a foreign government's health data. Newsquawk's framing is apt, that this reads as the chips-and-export-controls playbook applied to model access, treating pre-deployment testing itself as a chokepoint the US wants to control first.

What isn't yet answered: whether this becomes a standing US policy applied to every allied testing body, not just Britain's, and whether AISI gets any workaround before its next major model review. Neither the White House, Anthropic, nor OpenAI responded to Reuters' requests for comment as of Thursday.