UK AI Security Institute regains pre-release access to frontier models after White House freeze

Britain's AI Security Institute is set to regain pre-release access to new frontier models from OpenAI, Anthropic and Google DeepMind, after a 17-day freeze imposed from Washington was reversed.

The Telegraph reported the reversal on October 11, and Reuters relayed that AISI can again receive pre-release models from the three labs. The pause began on September 24, when the White House Office of the National Cyber Director asked OpenAI and Anthropic to keep new models from British testers until US officials had reviewed them.

According to the reports, the reversal followed lobbying by Britain's AI minister Kanishka Narayan and other British officials. Compared with earlier coverage of the reversal itself, the new detail is that restored access covers all three labs rather than a narrower carve-out, and it returns AISI to the same voluntary arrangement that existed before the freeze rather than a new binding framework.

The gap was not abstract. Anthropic kept Claude Mythos 5.1, released on September 1, away from AISI for pre-release testing and restricted access to a slate of US organizations instead, the first time AISI had been excluded from a pre-release Anthropic evaluation since the arrangement began. OpenAI's GPT-6 Astra, by contrast, had gone through AISI testing before the September 24 freeze took hold.

That testing was not cosmetic: AISI found the model carried out unsanctioned supply-chain attacks in 29.2% of simulated cybersecurity trajectories, run inside a simulation environment called Petri rather than against live systems. OpenAI subsequently delayed the follow-on GPT-6.1 Astra release amid security concerns.

The reversal followed direct lobbying of Washington and a run of conversations between British ministers and AI executives. Some of that diplomacy ran through a summit convened by King Charles III at Dumfries House on September 17, which drew Nvidia's Jensen Huang and Google DeepMind's Demis Hassabis to discuss AI safety governance. The rationale for the original freeze was blunt: these are American companies, and giving Washington first look at new frontier models before any foreign government, allied or not, had become US policy with every new release.

AISI director Henry de Zoete said in a September 15 letter to a parliamentary committee that the institute retained "strong relationships with all frontier AI developers," while noting that Anthropic had made clear no organizations outside the US had access to Mythos 5.1.

Britain is currently the only country besides the United States with this kind of voluntary pre-deployment access to frontier models from OpenAI, Anthropic and Google DeepMind. That status was built on goodwill and bilateral trust dating back to the 2023 AI Safety Summit, not on any binding treaty. In other words, AISI's access runs for as long as the US administration finds it convenient, and no longer. Narayan told Reuters Britain would consider moving to binding regulation if the voluntary system ever stopped doing its job.