US reverses course on Britain's testing of AI models, Telegraph says

The reversal concerns when US AI developers may share new models with Britain. The earlier request told developers such as Axiom and Paradox not to send models to Britain’s AI Security Institute before release. The administration has now withdrawn that restriction, according to the Telegraph report described by Reuters.

Early access means British evaluators can examine a model before ordinary users receive it. They can test what the model can do and assess possible misuse risks. This creates a review stage before public availability, rather than waiting until after launch.

The change gives US developers more room to cooperate with Britain. It also signals a softer Trump administration stance toward international AI safety work. The article does not describe a new formal testing agreement or say that every model will be shared. It reports that developers can share models without violating the administration’s position.

Britain’s AI Security Institute is a government-backed testing body created after the world’s first AI safety summit in 2023. Its role is to examine the most advanced AI models and assess both their capabilities and the risk that people could misuse them. The article presents it as a safety evaluator, not as an AI developer.

Receiving early access means the institute gets a model before its public release. Evaluators can run tests while developers still control the release process. They can identify important capabilities or misuse concerns before the model becomes broadly available. The article does not specify how much access evaluators receive or which tests they perform.

This timing matters because safety checks can happen before deployment rather than afterward. The reported policy reversal allows US developers to cooperate with the institute without conflicting with the Trump administration’s position. It therefore expands the institute’s potential role in international model testing.

Four key participants appear in the decision. The Trump administration first asked US developers not to share new models with Britain’s institute, then reversed that request. Axiom and Paradox are the developers specifically named in the report. Britain’s AI Security Institute is the external evaluator that may receive models early.

The US government’s role is setting the policy boundary for American developers. Axiom and Paradox build AI models and could provide unreleased systems for testing. The British institute tests advanced models, examining their capabilities and possible misuse risks. The article does not say that either company has already shared a particular model.

The policy change gives these US companies greater latitude to cooperate with Britain. It also makes the institute a channel for international safety work. Spokespersons for Trump, Axiom, and Paradox did not immediately respond to Reuters’ requests for comment, so their individual views on the reversal are not stated.

The article does not provide a numerical measure of how powerful these AI models are. It calls them “the most advanced AI models” and describes AI as increasingly powerful technology. That establishes their significance, but it does not give model sizes, performance scores, computing levels, or specific abilities.

The safety evaluation covers two broad areas. First, the institute tests model capabilities, meaning what the systems can do. Second, it examines the risk of misuse. The article does not list particular tasks, threats, or testing methods. Therefore, the safest description is broad rather than technical: evaluators study advanced models’ abilities and potential for harmful use.

Testing before release can help policymakers and developers understand a system before it reaches the public. The report connects this work with international AI safety cooperation. It also says the Trump administration’s task force requires developers to report security incidents, suggesting that oversight is considering both pre-release evaluation and problems discovered after deployment.

Allowing pre-release sharing could strengthen cooperation between American developers and British safety evaluators. Britain’s institute would be able to examine some US models before public release. That creates an international testing link at a stage when developers and policymakers are trying to understand powerful systems and misuse risks.

The key mechanism is permission. The earlier request discouraged companies from sharing new models with the institute. Reversing it removes that obstacle, giving developers such as Axiom and Paradox greater latitude. The institute can then test models, while developers can use the results to understand capabilities and possible risks before release. The article does not describe mandatory participation or a shared testing standard.

The immediate signal is a softer US position toward international AI safety efforts. This may support more cooperation between the two countries. However, the article does not establish how many models will be shared, what findings will be published, or whether other governments will adopt similar arrangements.

Britain’s AI Security Institute was set up after the world’s first AI safety summit in 2023. The article explains its practical purpose: testing the most advanced AI models for capabilities and misuse risks. It does not give a more detailed reason for the institute’s creation or describe the summit’s agreements.

The policy decision fits a broader regulatory debate. Trump has repeatedly dismissed calls for binding safety requirements on AI developers. The earlier request against sharing models with Britain reflected that tougher position toward international safety efforts. Reversing the request now permits more cooperation, although it does not create binding safety rules.

The article presents this as a possible shift in direction. Trump’s newly launched AI task force also said on Friday that developers must report security incidents involving their models. Together, these developments show attention moving toward selected oversight measures, while the article does not say that comprehensive regulation has been adopted.

An AI model is a trained computer system that uses patterns learned from data to produce outputs. Depending on its design, an output might be text, an answer, a prediction, or generated content. The source article does not explain this technical mechanism, so this description uses established AI knowledge. It focuses instead on models’ capabilities and misuse risks.

When a user provides an input, the model processes it and generates an output based on its learned patterns and instructions. Before release, safety evaluators can give the model challenging or harmful-use prompts and observe its responses. They can also examine whether it performs capabilities that developers or policymakers need to understand. The article does not identify specific tests.

Early testing may help reduce misuse by finding problems before broad public access. Developers and evaluators can then consider safeguards or release decisions. In this case, Britain’s institute tests advanced models, while the policy reversal gives US developers more freedom to provide models before release. The article does not promise that testing will prevent every harmful use.