Summary
- Reporting says US authorities asked OpenAI and Anthropic not to provide new models to the UK AI Security Institute before American review.
- AISI identifies pre-deployment access to leading models as one of the capabilities underpinning its government research work.
- The reported request would add another cross-border dependency to the UK’s ability to evaluate advanced models before wider deployment.
US authorities have reportedly asked OpenAI and Anthropic to delay providing new frontier models to Britain’s AI Security Institute until American officials have reviewed them, adding a new cross-border dependency to the UK’s model-testing programme.
Bloomberg reported on 25 September that the US administration had made the request to the two American AI companies, with a British official confirming the request to the news organisation.
The precise legal or administrative basis has not been publicly detailed. There is also no indication in the reporting that the arrangement constitutes a permanent prohibition on UK access to future models.
The development nevertheless intersects directly with the operating model of the AI Security Institute, the UK government research organisation within the Department for Science, Innovation and Technology.
AISI says pre-deployment access to leading AI models and close collaboration with developers give its researchers privileged access to development approaches and allow them to support safety-related decisions.
Its work includes technical research into risks posed by advanced AI systems rather than solely post-release policy analysis. Timely access is therefore important when capabilities and safeguards can change materially from one model generation to the next.
If US review becomes a prerequisite before British government researchers receive new models from American developers, the sequence of independent evaluation changes. The UK could continue to test the systems while becoming more dependent on decisions in Washington over when access is permitted.
The issue goes beyond institutional prestige. Governments are building technical capacity to assess whether advanced AI systems can automate cyber operations, circumvent safeguards, undertake longer autonomous tasks, or develop capabilities that materially alter security risk.
Those evaluations have greater value when they occur before or around deployment, while developers are still making release and mitigation decisions, rather than after a system has already entered broad use.
AISI’s existing model depends significantly on cooperation from frontier developers. Unlike conventional regulatory systems in which an authority may possess explicit statutory powers to demand product information or testing access, much frontier-model evaluation has been built through agreements and voluntary collaboration.
The reported US intervention exposes the geopolitical dependency inside that arrangement. Many of the most capable proprietary models are developed by US companies, meaning another government can potentially influence when foreign evaluators receive access even where the companies themselves have established working relationships with those evaluators.
The UK and US have also built extensive cooperation around AI security, making the reported restriction more complex than a simple dispute between competing national testing regimes.
AISI describes itself as a research organisation intended to equip governments with scientific understanding of advanced AI risks. Its ability to do that independently depends partly on whether researchers can inspect systems at the frontier rather than relying on delayed access, developer summaries, or public interfaces.
The reported request does not establish that OpenAI or Anthropic will permanently stop working with the institute, nor that every future model will be withheld. The unresolved issue is narrower: whether AISI can continue obtaining sufficiently early access to perform the pre-deployment evaluations it identifies as one of its core capabilities.
Until the US administration, the companies, or the UK government provide fuller details, the duration and scope of the request remain unclear.
