The White House has reportedly restored Britain’s access to advanced AI models for testing before release, giving researchers another opportunity to identify security risks before businesses adopt them.
In September, the White House asked OpenAI, Anthropic and other major AI developers to stop sharing new models with the UK’s AI Security Institute (AISI), according to The Telegraph. Following that intervention, Anthropic restricted British officials’ access to Claude Mythos 5.1.
The Telegraph reported that lobbying by UK AI minister Kanishka Narayan helped reverse the decision, restoring pre-release access to models from Anthropic, Google DeepMind and OpenAI. In an Oct. 12 speech, Narayan confirmed that AISI scientists were evaluating the latest generation of models, including Gemini 4 Argon, following discussions with developers and the US government.
Washington faces growing pressure over AI risks
The decision comes as governments grapple with how to evaluate increasingly capable AI systems without unnecessarily restricting development.
President Donald Trump has opposed tighter international controls on AI. Addressing the United Nations General Assembly last month, he said he wanted to “encourage” AI rather than “rein it in,” according to The Telegraph. Britain, meanwhile, has sought a larger role in international AI oversight. Its AISI, established in 2023, tests advanced systems to assess their capabilities and potential for misuse.
The reported reversal also follows a separate move by Trump’s newly launched AI task force demanding immediate reporting and remediation of incidents involving AI models, according to Axios. Together, the developments suggest that practical safety concerns are creating room for cooperation even as Washington resists binding restrictions on AI development.
UK considers stronger AI incident reporting
On Oct. 12, Narayan announced an expansion of the government’s use of advanced AI to strengthen cyber defenses. Ministers are also considering how the Cyber Security and Resilience Bill and the wider regulatory regime could better protect essential services from AI-related threats.
That work includes examining how serious AI incidents and near misses should be identified and reported. Details will be developed with parliamentarians, technical experts and affected operators.
The government is also developing plans with the National Cyber Security Centre, the Department for Digital, Culture, Media and Sport, and AISI to respond to serious incidents involving AI agents.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
A win for independent AI oversight
The reversal gives Britain more room to evaluate systems developed by some of the world’s leading AI companies. However, the episode exposes a weakness in relying on international cooperation, which political decisions can interrupt.
If independent evaluators lose access during critical development periods, dangerous capabilities could go unexamined until after a model reaches users. Restoring access helps, but its value will depend on whether developers can maintain that access consistently and whether testing findings lead to meaningful security improvements.
The UK also wants to use its G20 presidency in 2027 to encourage international cooperation on shared AI principles and standards, according to The Telegraph. That ambition will be harder to achieve if access to advanced systems depends on shifting diplomatic priorities.
What the decision means for AI users
For businesses choosing AI tools, the practical value of pre-release testing is additional scrutiny of capabilities and safeguards before deployment. IT teams should look for published findings that explain what researchers tested, what weaknesses they identified and how developers responded.
Access alone does not establish that a model is suitable for a particular organization. A model tested by government researchers can still behave differently when connected to company documents, internal applications or tools that let it take action.
Before deploying AI in sensitive workflows, teams should assess the data and systems it can access, limit permissions to the task, and decide which actions require human approval. Monitoring should also capture enough information to investigate unexpected behavior and support incident response.
UK organizations supporting essential services should watch how the proposed reporting measures develop, including which incidents and near misses would fall within scope. For now, the announcement signals possible changes rather than settled new AI reporting obligations.
Read more: Europe has faced similar challenges securing independent testing access, with EU officials gaining access to Anthropic’s Mythos 5 while its newer model remained unavailable.