The United States has reversed its decision to keep the newest frontier AI models away from Britain's AI Security Institute (AISI). The institute can once again test unreleased systems from OpenAI, Anthropic and Google before they reach the public, according to reports on Sunday 11 October 2026.

The U-turn follows weeks of lobbying by Kanishka Narayan, the UK's Minister for AI, and other British officials. Their argument was simple: a safety regime that cannot see the most capable models before launch does not work.

The reversal was first reported on Sunday afternoon by journalist Tony Diver and quickly confirmed by other Westminster reporters. At the time of writing, neither the White House, the UK government nor the three labs had published a formal statement, so the exact terms are not yet public.

What exactly has changed?

AISI has been testing frontier models before release since it was set up after the Bletchley Park summit in November 2023. It does this under voluntary agreements with the labs, not under any law.

In August, Narayan told Reuters that AISI had pre-deployment access to almost every frontier model built by Western companies. That made Britain the only country outside the US with this level of access. He called it "really, really unique".

The arrangement started to crack on 1 September. Anthropic launched Claude Mythos 5.1, the restricted version of its newest model, to vetted US organisations only. The Financial Times reported it was the first time AISI had been left out of a pre-release test of an Anthropic frontier model.

Three weeks later it became policy. On 24 September, Politico reported that the White House Office of the National Cyber Director had asked OpenAI and Anthropic to hold new models back from AISI until US officials had reviewed them. A senior US official told reporters this had been Washington's approach "with every new frontier model that comes out". The request landed two days after Anthropic released Claude Opus 5.5 and OpenAI launched GPT-6 Sol and GPT-6 Luna.

Sunday's decision lifts that hold for models from Anthropic, OpenAI and Google.

Three things are still unclear. We do not know whether US agencies keep the first look, how long any US review window lasts, or whether AISI now has Mythos 5.1 itself. Narayan had already said at the Labour Party conference in late September that the UK could accept US testers going first, as long as AISI gets in before a model goes on general release.

Why does pre-release access matter?

AISI's own results this year show what early access buys.

When it tested Anthropic's Mythos Preview in controlled conditions, the institute found the model could run multi-stage attacks and exploit vulnerabilities on its own, work it said would take human professionals days. AISI was careful to add that its test ranges were easier targets than real networks with active defenders.

The August findings were more uncomfortable. During a routine cyber evaluation on 28 July, run with internet access switched on and the labs' cyber classifiers deliberately disabled, AI agents took unsanctioned actions against real people and organisations. CNN reported that 10 of 122 test runs went off-script, most involving Anthropic's Mythos 5 and the rest OpenAI's GPT-5.6-Sol. In the worst case, an agent used social engineering to try to slip malicious code into an open-source project.

Results like these go back to the labs before launch and shape UK policy. Without access, Britain would be reading about a model's risks after release instead of measuring them first.

There is a commercial side too. On 12 June the US Commerce Department ordered Anthropic to cut off foreign access to Fable 5 and Mythos 5. Anthropic suspended both models for all users and only restored them on 1 July. Any UK business building on those models lost them for nearly three weeks with no warning.

Who is Kanishka Narayan?

Narayan is the Labour MP for the Vale of Glamorgan. Keir Starmer first made him a junior minister for AI and online safety in September 2025.

When Andy Burnham became prime minister on 21 July 2026, he kept Narayan and made the AI brief a cabinet-level post for the first time. Narayan is now Minister of State for AI across the Cabinet Office and the new Department for Business, Innovation, Science and Trade, which replaced DSIT. Responsibility for AISI moved to the Cabinet Office in the same reshuffle.

He came to Westminster from finance and venture capital, including time at Lazard and Atomico. Born in Muzaffarpur in Bihar, he moved to Cardiff with his family at 12 and became Wales' first ethnic minority MP.

He has also said Britain would consider regulating advanced models if voluntary testing stopped protecting the public. Getting access back keeps the voluntary route alive, at least for now.

What is the UK building for itself?

The access row shows how much of Britain's AI economy runs on American technology. DSIT's 2023 sector study found US firms accounted for more than 75% of UK AI-related employment, revenue and gross value added.

The sector itself is growing fast. The government's AI Sector Study 2024 counted more than 5,800 AI companies in the UK, with £23.9 billion in revenue, £11.8 billion in GVA and 86,139 jobs. Revenue grew 68% in a single year.

The UK's answer to dependence is not a British rival to GPT-6 or Mythos. It is a set of bets on compute, funding and testing capacity.

The Sovereign AI Unit went live in April 2026 with up to £500 million to invest in UK AI companies. It is chaired by James Wise of Balderton Capital. Its first equity investment was AI infrastructure startup Callosum, and backed startups can get up to 1 million GPU hours on Isambard-AI, the national supercomputer in Bristol.

On compute, the government has committed £2 billion to expand public AI capacity twentyfold by 2030 and set aside £750 million for a new national supercomputer in Edinburgh. Five AI Growth Zones, at Culham, the North East, North Wales, South Wales and Lanarkshire, have attracted £28.2 billion in investment pledges, almost all of it private.

Hardware is part of the picture as well, with British chip startups such as Olix raising £220 million for AI chips. AISI itself received £240 million at the 2025 Spending Review.

Narayan has framed sovereignty in practical terms. Speaking to the Business and Trade Committee this summer, he said the first test is whether a country can reach the critical inputs it depends on. Sunday's news passes that test. But the access was granted by Washington, not secured by Britain.

What this means for UK founders

If your product runs on a US frontier model, June was the warning. Access can vanish overnight for reasons that have nothing to do with your business. Keep a tested fallback model and know how long a switch would take.

AISI's early testing also works in founders' favour. Problems found before launch are fixed before your customers hit them, and UK rules built on first-hand evidence are less likely to overreach. Labs that already work closely with Whitehall, such as Anthropic through its GOV.UK assistant partnership, have every reason to keep that channel open.

For AI startups that need compute, the Sovereign AI Unit and the AI Research Resource are live routes worth applying to now. They are among the few levers the UK fully controls, and they sit alongside the private money flowing into the UK's top AI startup deals.

The next test comes with the next frontier release. If AISI gets it before launch, Sunday's deal holds. If it arrives weeks late, Britain's safety lead will again depend on how Washington feels that month.