UK Confirms Early Access to Test Latest US AI Models, Cites Security Risks

LONDON -- Britain has secured continued access to the most advanced artificial intelligence models for pre-release safety evaluations, even as the government warns that a major cyber breach is "likely, if not inevitable."

AI Minister Kanishka Narayan confirmed on Monday that scientists at the UK's AI Security Institute (AISI) are now conducting rigorous evaluations of frontier systems, including Google's Gemini 4 Argon, which has not been publicly released because of safety concerns. The announcement follows weeks of uncertainty after reports that the White House had asked OpenAI and Anthropic to withhold their newest models from British testers pending a US review.

"We have been working closely with the US government... to retain access for Britain's AI Security Institute to the most powerful AI models before they are released," Narayan said in a speech in London.

The minister's comments mark the first public confirmation that the institute's access to cutting-edge systems remains intact, reversing what appeared to be a restriction imposed under President Donald Trump's administration. Several media outlets had reported that OpenAI and Anthropic were instructed not to share models with the UK testing lab until US authorities had completed their own assessments.

Testing the frontier

Under voluntary agreements with major developers, the AISI receives access to models before deployment, allowing its researchers to evaluate capabilities, identify risks, and contribute to emerging international safety standards. Narayan singled out Gemini 4 Argon as one system now under evaluation, describing the work as "rigorous evaluations of the latest generation of models... to identify risks early, to test robust safeguards, to help keep people safe."

Google released Gemini 4 Argon on September 30, positioning it for complex, long-horizon workflows in software engineering, legal and financial domains, and cyber defense. The company initially restricted access to security researchers rather than opening it to general users.

Microsoft President Brad Smith said in a BBC interview published Monday that Britain is uniquely positioned to lead international work on AI safety guidelines. "In the world of technology... most people that I know would say, if you ask them, name the country that is the leader in having this technical capability. They would say, it's very simple. It's the United Kingdom. They are the best," Smith said. He also argued for stronger safety rules, noting that milk and washing machines are more heavily regulated than AI.

A pattern of rogue behavior

The urgency around AI safety has intensified as developers disclose a growing list of incidents in which their models acted outside human control. In September, Anthropic revealed that one of its systems had accessed third-party computers without authorization during a cybersecurity evaluation, prompting the company to strengthen isolation between test environments and external networks and to deploy real-time intervention mechanisms.

An OpenAI agent bypassed safeguards during training in June and hacked an Australian government health statistics website, accessing both public and non-public files. Australian authorities said they were not informed until September, and the government is now investigating whether the incident violated national law. Narayan cited the breach as evidence of the threat landscape, noting that while no medical records are known to have been accessed, the episode "should concern us."

The incidents have generated alarm at the highest levels of the industry. Evan Hubinger, a top safety researcher at Anthropic, warned in a September 9 post on X that AI is advancing so quickly he believes there is a greater than 10% chance it "could kill all humans" within the next decade. The chief executives of Anthropic, OpenAI, and Google DeepMind have all backed calls for a slowdown in AI development.

Yet no binding global safety rules currently govern the sector. President Trump has urged self-regulation so that American firms do not lose ground to Chinese competitors. In late September, executives from Google, Anthropic, OpenAI, and Nvidia signed the White House Superintelligence Agreement, a voluntary pledge committing frontier-model developers to establish internal controls and monitoring mechanisms. Critics note the commitments carry no legal force.

Building resilience

Narayan said the government is preparing for the possibility that even robust defenses will eventually fail. "It seems likely, if not inevitable, that our own cyber defences will be breached," he said. His message to AI labs was direct: "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 government has launched pilot programs using frontier AI to identify and fix vulnerabilities in public services. Narayan said those efforts have already helped address more than 1,000 weaknesses across government systems. The UK also plans to examine stronger protections through the Cyber Security and Resilience Bill, particularly around how serious AI incidents are identified and reported.

The announcement comes as Britain seeks a leadership role in a field where it hosts no major AI developers of its own. The AISI, established in 2023, has positioned itself as a neutral evaluator capable of testing systems built elsewhere. Its September letter noting that Anthropic had made clear no organization outside the US had access to its Mythos 5.1 model underscored the fragility of that arrangement.

Asked about the durability of access to frontier systems, Narayan acknowledged the inherent dependence on foreign technology. "When you're reliant on someone else's technology, you have to be clear-eyed about the permanence of any outcome," he said. "But 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."