The Trump administration has reversed an earlier request that American artificial intelligence developers withhold new models from Britain’s AI Security Institute before public release, allowing companies such as OpenAI and Anthropic to provide the UK government with early access for safety testing, the Telegraph reported on Sunday.
The reported reversal marks a shift in Washington’s approach to international AI safety cooperation, giving US developers greater latitude to work with British researchers as governments seek to assess the risks posed by increasingly capable AI systems.
The newspaper did not provide further details about the revised guidance.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
The development comes amid an ongoing debate over how governments should oversee advanced AI without unnecessarily restricting technological development. Early access allows safety researchers to examine a model’s capabilities and potential vulnerabilities before it becomes widely available, when weaknesses could be more difficult to contain.
Britain’s AI Security Institute was established following the world’s first AI Safety Summit in 2023. It tests advanced AI models to evaluate their capabilities and the risks of misuse, providing policymakers with technical assessments as they consider how to respond to the technology’s rapid development.
Allowing American developers to share unreleased models with the institute could strengthen this work by giving researchers an opportunity to examine systems before their capabilities and limitations become publicly known. However, the extent of the cooperation permitted under the reported policy change remains unclear.
The reversal also comes as the Trump administration’s position on AI oversight faces renewed attention. Trump has repeatedly dismissed calls for binding safety requirements on AI developers, favoring an approach that has placed greater emphasis on technological development than on imposing extensive regulatory obligations.
Yet the administration has recently signaled that certain security measures remain necessary. Trump’s newly launched AI task force said on Friday that developers must report security incidents involving their AI models, suggesting that the administration is considering specific obligations even as it resists broader mandatory safety requirements.
The incident-reporting requirement and the reported reversal on UK model access point to a more nuanced policy approach than a blanket rejection of AI safeguards. They do not, however, establish that Washington has embraced comprehensive safety regulation.
The ability to examine AI models before release is necessary because some risks may not be apparent from public demonstrations or routine testing. Researchers can conduct controlled evaluations to determine whether advanced systems can assist malicious activities, expose vulnerabilities, or behave in ways that create security concerns.
The UK institute’s work is part of a broader effort to build technical capacity for evaluating increasingly capable models. By testing systems before deployment, researchers can potentially identify weaknesses early enough for developers to address them before wider use.
Cooperation with American AI companies is deemed necessary now because US developers include some of the leading firms building advanced models. Access to their unreleased systems would give British researchers more opportunities to evaluate emerging capabilities rather than relying exclusively on publicly available versions.
The reported change also highlights the international dimension of AI safety. Models developed in one country can be distributed globally, while security risks may affect users, businesses and public institutions across borders. Cooperation between governments and developers can therefore help researchers compare findings and assess threats that are not confined to a single jurisdiction.
Nevertheless, early access alone does not guarantee that identified risks will be resolved. Its effectiveness depends on the scope of testing, the information developers provide, the willingness to address findings, and the safeguards governing the handling of sensitive model information.
Safety Policy Remains A Point of Tension in Washington
The reported reversal comes against the backdrop of Trump’s resistance to binding AI safety requirements. His administration has generally prioritized the development and deployment of AI, while policymakers and researchers continue to debate how much oversight is necessary to manage risks without slowing innovation.
The requirement announced by the administration’s AI task force that developers report security incidents introduces a specific accountability measure into that debate. Incident reporting can help authorities identify emerging threats and understand how AI systems are being exploited or failing in practice.
However, reporting obligations differ from rules requiring developers to submit models for independent evaluation, demonstrate compliance with safety standards, or delay releases until identified risks are addressed. The reported decision to permit cooperation with Britain’s institute should therefore not be interpreted as evidence that the administration has adopted a broader system of mandatory pre-release testing.
The change is expected to make it easier for American AI companies to cooperate with a foreign government research body without risking a conflict with the administration’s earlier position. For Britain, access to unreleased models could strengthen its ability to evaluate advanced systems before they reach users.