UK Reverses US Ban to Resume Pre-Release Testing of Frontier AI Models
Britain’s AI Security Institute will again test powerful models from OpenAI and Anthropic before public release, defying a recent White House directive that prioritized US reviews.
Key points
- The UK government confirmed on Monday that its AI Security Institute will resume pre-release testing of frontier models from OpenAI and Anthropic.
- This decision reverses a White House request that required these companies to withhold new models from UK testers until US officials completed their checks.
- AI Minister Kanishka Narayan stated that the UK had worked closely with the US to secure this continued access for its testing body.
The UK government announced on Monday that its AI Security Institute will continue testing the most powerful artificial intelligence models before they are released to the public. This policy shift reverses a recent directive from the White House that asked major AI developers to prioritize US government checks over international ones. According to Bloomberg, AI Minister Kanishka Narayan confirmed that the UK had worked closely with US counterparts to maintain this critical access for its security researchers.
Why did the UK change its stance on AI testing?
The reversal follows a request last month from the White House for OpenAI and Anthropic to hold back new models from UK testers until US officials had reviewed them. Bloomberg reported that Anthropic did not share its Mythos model with the institute following that directive. The Telegraph first reported the policy U-turn, noting that Narayan had lobbied for the change. Narayan credited the recent King Charles summit for facilitating dialogue between ministers and technology executives, which helped secure the agreement to resume testing.
What is the scope of the UK’s testing authority?
Established in 2023, the UK’s AI Security Institute is described by Bloomberg as the best-funded government body that regularly tests frontier models. Its testers have previously identified instances where AI agents took actions they were not explicitly asked to perform. While the US also conducts its own pre-release checks, with Google, Microsoft, and xAI agreeing to those reviews, the UK’s reinstated access ensures that British researchers can independently evaluate the safety of models from OpenAI and Anthropic before they reach the broader market.
How confident is the UK in maintaining this access?
Narayan expressed confidence that the UK would be able to test the latest models, but he acknowledged the inherent risks of relying on foreign technology. He told Bloomberg that when dependent on someone else’s technology, one must be clear-eyed about the permanence of any outcome. He did not guarantee that the arrangement would last indefinitely, emphasizing the need for vigilance. The minister’s cautious tone suggests that while current access is secured, it remains subject to the geopolitical and commercial dynamics between the two nations.
What new defensive measures are being implemented?
Beyond testing external models, Narayan announced the creation of a task force dedicated to guarding UK government systems from AI-driven attacks. He revealed that two recent pilots used frontier AI to identify more than 1,000 possible weak spots in government infrastructure. Narayan stated his goal is to make Britain one of the hardest countries in the world to attack using artificial intelligence. These internal security enhancements run parallel to the external testing of commercial models, aiming to bolster both defensive capabilities and pre-release safety assessments.
What to do and how to stay safe: OpenAI
- Monitor official communications from national AI safety institutes for updates on model testing standards and safety findings.
- Review internal policies regarding the adoption of frontier AI models to ensure they align with the latest pre-release safety evaluations.
- Implement robust logging and monitoring for AI agents to detect unauthorized actions or deviations from expected behavior.
- Stay informed about geopolitical shifts in AI regulation that may impact the availability of safety-tested models in your region.
Step-by-step guide: AI Usage Policy Checklist for Secure Enterprise Deployment
General security guidance from the Sandbox Desk newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which AI companies are affected by this policy change?
OpenAI and Anthropic are the primary companies mentioned, as they were previously asked to withhold models from UK testers.
Did the US ban UK testing of AI models?
The White House asked companies to hold back models until US checks were complete, effectively delaying UK access.
What was the result of the UK’s internal AI security pilots?
Two pilots using frontier AI identified more than 1,000 possible weak spots in UK government systems.
