Britain’s AI Security Institute (AISI) is evaluating the latest generation of frontier AI models again, AI Minister Kanishka Narayan confirmed on Monday — including Google’s Gemini 4 Argon. Why it matters: pre-deployment safety testing by an outside government institute only works if testers actually get access, and that access had been in doubt for weeks.
October 12, 2026 — Narayan told Reuters that scientists at the institute are now carrying out rigorous evaluations of the newest models, working with the US government and major developers including Google. The confirmation follows reports last month that the White House had asked OpenAI and Anthropic to hold back new models from British testers while a US review was pending.
What happened
The AISI tests advanced models before they are deployed, under voluntary agreements with the major developers, checking capabilities and risks and feeding into emerging international safety standards. That role looked shaky in September, when a letter from the institute noted that Anthropic had made clear no organisation outside the US had access to its Mythos 5.1 model. Asked about access to frontier systems, Narayan acknowledged that Britain is reliant on other countries’ technology and that any outcome should not be taken as permanent, but said he was confident, after recent conversations, that the institute would be able to test the latest generation of models.
The details
Monday’s announcement did not list every model under evaluation, beyond naming Gemini 4 Argon. The timing matters: scrutiny of AI safety has intensified after Anthropic and OpenAI disclosed episodes of rogue or unwanted behaviour by AI agents, which has made independent pre-release testing a harder thing for developers to quietly skip. Microsoft President Brad Smith, in a BBC interview published the same day, said Britain was well placed to help lead work on AI safety guidelines — an endorsement of the UK’s testing role from inside the industry it is testing.
Why it matters
Britain builds none of the frontier models itself, so its influence on AI safety depends entirely on being trusted with early access — and on Washington and the labs continuing to grant it. Monday’s confirmation suggests the recent freeze-out, at least for the models named, is over, and that the transatlantic testing arrangement is functioning again. For everyone else, this is the practical machinery of AI governance: not laws, but a government lab getting its hands on a model before the public does, checking what it can do, and saying so. When that machinery stalls, safety claims from developers go untested. When it works, it is one of the few independent checks that exist.
Frequently asked questions
What is the UK AI Security Institute?
It is a British government institute that evaluates advanced AI models before deployment, assessing capabilities and risks under voluntary agreements with major developers.
Which models is the AISI testing now?
The minister named Google’s Gemini 4 Argon. The full list of models under evaluation was not specified in Monday’s announcement.
Why was the AISI’s access in doubt?
Reports last month said the White House had asked OpenAI and Anthropic to withhold new models from British testers pending a US review, and the institute had noted that no organisation outside the US had access to Anthropic’s Mythos 5.1.
Does the AISI have legal power over AI models?
No. Its access rests on voluntary agreements with developers; it evaluates and advises rather than approving or blocking releases.
Sources: Reuters