UK minister says AISI is testing the latest models, including Gemini 4 Argon

Kanishka Narayan, the AI minister, told an audience at Imperial College London on 12 October 2026 that Britain has been working with the US government and AI labs to keep pre-release access for the AI Security Institute, and that institute scientists are now evaluating current models. The speech was published at 12:59:53+01:00 (11:59:53 UTC). The line is:

That is why, over the last few weeks, we’ve been working closely with the US government, the labs - from His Majesty’s gathering of leading AI developers at Dumfries House in Scotland to discussions at the United Nations General Assembly in New York - to retain access for Britain’s AI Security Institute to the most powerful AI models before they are released. And I can confirm that AI scientists at the AI Security Institute are now conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon to identify risks early, to test robust safeguards, to help keep people safe.

The speech names Gemini 4 Argon. It does not name an OpenAI or Anthropic model as one of the systems under evaluation. Google announced Gemini 4 Argon on 30 September, with a phased release that Google described as entering a US voluntary pre-release process. The minister’s sentence does not say whether the copy AISI is evaluating is a pre-release build.

The same speech’s message to labs is: “Listen when your scientists warn of severe dangers. Investigate unexpected behaviours. And if you can’t adequately control the risks, do not release the system.” The press release published the same day uses different words in that sentence: “serious dangers,” “unexpected behaviour,” and “if you cannot adequately control the risks, do not release the system.” Tech.eu reports the press-release wording.

What is reported about the United States

The government pages say the work of recent weeks was with the US government and the labs. They do not describe a US request to withhold models, and they do not say that request was withdrawn. No statement from the US government, or from the Office of the National Cyber Director, was among the pages opened for this story.

Quartz, in a 12 October story by Cris Tolomia, reports a reversal and attributes the underlying account to other outlets. Quartz writes that the Trump administration told OpenAI and Anthropic to withhold models from the AI Security Institute until US authorities cleared them, according to Bloomberg. Quartz writes that this led Anthropic to stop British officials from reviewing Claude Mythos 5.1, according to LBC, and that Anthropic gave no explanation, according to Bloomberg. Quartz writes that The Telegraph first reported the reversal, according to LBC. Bloomberg’s own story was not opened. It is paywalled from here, so this account does not cite Bloomberg as read. LBC and The Telegraph were not opened either. The withholding, the Mythos restriction, and the reversal are reports of reports.

Quartz quotes Narayan speaking to Bloomberg: “I’m very confident that, as a consequence of the conversations we’ve had, we will be able to test the latest generation of models.” And: “When you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome.” Those sentences are Narayan’s as Quartz relays Bloomberg. They are not in the gov.uk speech text.

Pilots, the bill, and a preview

The press release says two pilots “have already helped identify and fix over a thousand potential cyber vulnerabilities across government systems.” The speech says that across two pilots, frontier AI “has helped identify and address more than a thousand vulnerabilities, alongside other threats,” and that a central vulnerability-monitoring service has helped cut the time to fix identified weaknesses from 50 days to eight. Those figures are the government’s.

Both the speech and the press release name the Cyber Security and Resilience Bill. The press release says the government will explore strengthening parts of that bill and the wider regime, including how serious AI incidents and near misses are identified and reported, and “AI loss of control risk, where autonomous AI systems evade human oversight.” It says work with the National Cyber Security Centre and the AI Security Institute is looking at how to respond to incidents involving AI agents. Tech.eu’s story uses the plural “Bills.” The gov.uk pages use the singular bill.

The Guardian published a preview the same day, written in the future tense (“will say”). It quotes Geoffrey Irving, a former chief scientist at the AI Security Institute: “The AI companies do not have security that would prevent an adversarial country from stealing the models.” The Guardian says OpenAI says it has multiple protections against theft of its models, including encryption and continuous monitoring. That exchange is Irving’s warning and OpenAI’s reply, as the Guardian reports them. It is not part of the published speech.