The UK’s AI Security Institute (AISI) is set to regain early access to new models from Anthropic, Google DeepMind, and OpenAI. The White House reversed its decision to block that access after diplomatic pressure from AI Minister Kanishka Narayan, according to The Daily Telegraph.
Access restored after a September block
In September, the Trump administration told the companies to stop sharing their models with AISI. Anthropic then restricted the institute’s access to its new Claude Mythos 5.1 model.
The latest reversal means the institute will again be able to examine frontier models early. That access matters because AISI tests AI systems for safety risks before they ship, giving it a chance to assess potential concerns ahead of broader release.
The source does not specify when access will resume or which particular models will be available beyond naming the three companies. It reports that the block has been lifted.
Safety testing and incident reporting
Narayan plans to announce new AI safety measures in an upcoming speech. The plans include stricter reporting requirements for AI-related security incidents involving critical infrastructure.
Those reporting requirements would focus on incidents connected to critical infrastructure. The source does not provide further detail about the proposed rules, their scope, or how they would be implemented.
Together, the renewed model access and planned reporting measures point to two parts of AI safety work: examining systems before they are released and receiving information about security incidents after they arise. The article describes the institute’s pre-release testing role and Narayan’s planned reporting requirements, without setting out a broader policy framework.
Different views on international AI oversight
The reversal comes amid disagreement over how much international cooperation should shape AI development. Prime Minister Andy Burnham had previously pushed for international AI regulation at the UN General Assembly.
At the same event, Trump warned that “globalist” organizations should stay out of American AI development. The source presents those remarks alongside the earlier decision to block model sharing and the subsequent diplomatic effort to restore AISI’s access.
The contrasting positions show why access to frontier models can become a diplomatic issue as well as a technical one. A safety institute can only test models made available to it, while decisions about sharing those models may reflect wider disagreements over who should oversee AI development.
Security concerns remain part of the debate
UK security leaders have also raised concerns about the risks associated with AI. The heads of MI6, MI5, and GCHQ warned ministers that AI was empowering hostile actors and terrorists.
The report does not link that warning directly to the restored access or to the proposed incident-reporting requirements. It does, however, place the access decision in a wider context: governments are weighing how to evaluate new AI systems while responding to security concerns.
For AISI, restored early access gives the institute the opportunity to resume testing models from the named developers before they ship. Narayan’s planned announcement adds a separate focus on reporting security incidents involving critical infrastructure. The specific details of those measures remain to be announced.