White House reverses course and restores UK AI Safety Institute's access to frontier models

Access to cutting-edge AI models for UK safety testing is now restored, enabling earlier identification of risks in new AI systems.

The White House has reversed a prior decision, restoring the UK's AI Safety Institute (AISI) access to advanced AI models. Initially, the Trump administration had directed AI companies like Anthropic, Google DeepMind, and OpenAI to cease sharing their frontier models with AISI. This move by the US in September led to Anthropic restricting access to its new Claude Mythos 5.1 model. Following diplomatic efforts, particularly from UK AI Minister Kanishka Narayan, this restriction has now been lifted. Narayan is expected to announce new AI safety measures, including enhanced reporting for AI-related security incidents affecting critical infrastructure. The diplomatic push for AI safety and regulation has been a growing theme, with UK Prime Minister Andy Burnham advocating for international AI governance at the UN General Assembly, contrasting with Trump's warnings against "globalist" interference in US AI development.

This decision signals a significant thaw in US-UK diplomatic relations concerning AI safety, moving from a stance of restriction to cooperation. It suggests a growing consensus among key global players that early access and collaborative safety testing are crucial for mitigating potential risks associated with increasingly powerful AI models, especially as concerns mount about AI empowering hostile actors. The UK's renewed access will likely lead to more proactive identification and addressing of safety vulnerabilities before models are widely deployed.

The restoration of access will allow AISI to continue its critical role in evaluating AI systems for safety before their public release, particularly concerning new iterations from major AI labs. This renewed collaboration is vital for international AI governance and standard-setting, indicating a pragmatic approach to AI safety that prioritizes information sharing over isolation. It may also influence other nations to adopt similar cooperative frameworks for AI risk assessment and management.