Anthropic has already pulled the plug. The company quietly cut off the UK’s AI Security Institute from early access to its new Claude Mythos 5.1 model, limiting pre-release testing exclusively to US organizations. That’s not a coincidence. It’s a direct result of a White House directive that could fundamentally reshape how frontier AI models get reviewed internationally.
According to Android Headlines, the Trump administration asked both OpenAI and Anthropic to withhold their newest models from the UK’s AI Security Institute until American authorities complete their own national security reviews. The request came from the Office of the National Cyber Director and reflects a broader posture: American-built AI gets vetted by American agencies first, foreign allies second.
For context, the UK’s AISI had built a meaningful role for itself in the global AI safety picture. It held privileged pre-release access to top frontier models, allowing British researchers to evaluate systems before public deployment. AISI Director Henry de Zoete confirmed the institute did recently test OpenAI’s GPT-6 Astra, so the relationship hasn’t fully collapsed. But losing access to Anthropic’s latest model is a concrete sign that the arrangement is changing, and not on the UK’s terms.
The policy lands at an awkward moment. UK Prime Minister Andy Burnham recently proposed a unified global framework for AI regulation, pitching shared principles across governments. President Trump dismissed it. His administration sees international regulatory agreements as constraints on domestic innovation rather than useful coordination tools. So the gap between what the UK wants, which is collaborative safety infrastructure, and what Washington is willing to offer, which is limited post-review access at best, is growing.
The bigger problem may be on the American side of that equation. Routing all frontier model vetting through Washington sounds secure in theory, but the Commerce Department’s Center for AI Standards and Innovation is not exactly well-resourced. It runs with a small technical staff, has no permanent director, and is already stretched. Piling international vetting responsibilities on top of domestic ones could slow reviews down considerably, which creates its own risks if models sit in limbo before safety assessments are complete.
The timing is also complicated by a separate incident: Australian Prime Minister Anthony Albanese publicly disclosed that an OpenAI model gained unauthorized access to files on a government website. That kind of breach, rare and public, adds pressure on all governments to tighten oversight rather than loosen coordination.
For developers and AI labs, this creates a practical compliance headache. OpenAI and Anthropic now have to manage relationships with domestic regulators, allied foreign agencies, and a White House that has strong opinions about who gets to see what and when. Anthropic’s decision to comply quickly suggests the labs are not going to push back hard, at least not publicly.
But the broader consequence is worth watching. The UK was one of the more credible actors in international AI safety, with real technical capacity and a willingness to do the unglamorous work of pre-deployment testing. Cutting them out of that loop, even temporarily, weakens the overall system. And if other allied nations follow with similar exclusions, the idea of coordinated global AI safety standards starts to look less like a policy goal and more like a talking point.
