UK Says Its AI Security Institute Will Keep Testing Frontier AI Models Before Release

UK AI Minister Kanishka Narayan said on October 12, 2026 that the UK AI Security Institute has kept access to the most powerful AI models before they are released, after weeks of work with the US government and the AI labs. He said the institute's scientists are now running evaluations of the latest generation of models, including Google's Gemini 4 Argon.

The announcement follows a standoff in transatlantic AI policy and safety that began in September, when the White House asked OpenAI and Anthropic to hold new models back from British testers until US officials had checked them. The institute is the body that Technology Secretary Peter Kyle renamed from the AI Safety Institute on February 14, 2025.

What did the UK government announce?

The UK government said on October 12, 2026 that scientists at the AI Security Institute are conducting "rigorous evaluations of the latest generation of models," following work with leading AI developers and the US government. Narayan made the announcement in a speech at Base One, Imperial College London's science and tech incubator.

In the speech, Narayan said the government had worked with the US and the labs over recent weeks, from King Charles III's gathering of AI developers at Dumfries House in Scotland to talks at the UN General Assembly in New York, to retain that access.

He also addressed the labs directly:

The model access news came inside a wider package, set out in a government press release. Two government pilots used frontier AI to identify and address more than 1,000 potential cyber vulnerabilities, and a central monitoring service cut the time to fix identified weaknesses from 50 days to eight, according to the speech. Ministers will extend that work across government systems, consider adding AI incident and near miss reporting to the Cyber Security and Resilience Bill, and work with the National Cyber Security Centre and the institute on how to respond to incidents involving AI agents.

What did the White House ask for in September?

The White House asked OpenAI and Anthropic not to share new AI models with the UK institute until the US government had tested them, according to a Politico report published on September 24, 2026 (Politico via The Next Web). According to that report, the request came from the Office of the National Cyber Director, and a senior administration official said the US wanted to review new models first.

Anthropic had already held back. It did not give Claude Mythos 5.1, released on September 1, to the institute before launch, the first time a major model was kept from it (FT via The Next Web). The institute's director, Henry de Zoete, acknowledged the missing access in a September letter to a parliamentary committee and said the institute had tested OpenAI's GPT-6 Astra before its release, The Next Web reported.

Did the White House reverse its decision?

The White House has reversed its decision to block the institute's access, according to a Telegraph report published on October 11, 2026 that also says Narayan lobbied for the change (Telegraph via The Decoder). The Decoder writes that the institute will again get early access to models from Anthropic, Google DeepMind and OpenAI, citing the Telegraph report.

The UK government's own speech and press release frame it differently. They describe work to "retain access" and do not mention a White House decision, OpenAI or Anthropic by name. Narayan was cautious about how long the arrangement will last, according to Bloomberg (Bloomberg via The Next Web):

He also said he was "very confident" the UK would be able to test the latest models (Bloomberg via The Next Web).

Which models is the institute testing now?

The only model the UK government named is Google's Gemini 4 Argon, from a company that the September request, as reported, did not cover. Google announced Argon on September 30, 2026 as the flagship of its Gemini 4 lineup, gave it to select cybersecurity partners first and set no public release date, Reuters reported (Reuters via The Korea Times). Google also said it takes part in the US government's voluntary pre-release review process.

That means the one confirmed evaluation involves a model that is not yet generally available, which fits the "before release" claim, but it does not show whether Mythos 5.1 or a future OpenAI model has reached the institute. In the US, the Commerce Department's Center for AI Standards and Innovation has had pre-release access to all five major US frontier labs since Google, Microsoft and xAI joined on May 5, 2026, according to The Next Web.

What we don't know yet

• Whether the White House has confirmed the change, and on what terms, in a public statement of its own.
• Whether Anthropic has now given the institute access to Claude Mythos 5.1, which it withheld in September.
• Whether OpenAI and Anthropic models released from now on will reach the institute before or only after a US review.
• Which models besides Gemini 4 Argon the institute is evaluating now.