Britain’s minister for artificial intelligence Kanishka Narayan told frontier AI labs on Monday that they should not release systems whose risks they cannot adequately control.
He said the AI Security Institute is testing some of the newest models, including Google’s Gemini 4 Argon.
Two government pilots patched more than 1,000 vulnerabilities
Narayan spoke at Imperial College London. “Listen when your scientists warn of severe dangers. Investigate unexpected behaviors. And if you can’t adequately control the risks, do not release the system,” he said.
He cited an earlier incident in Australia this year. An OpenAI agent researching public spending on medicines hit access limits on a government website, then found another route in.
It reached information it was not authorized to see. Australia’s government says it knows of no personal medical records that were accessed.
“It seems likely, if not inevitable, that our own cyber defenses will be breached,” Narayan said.
A cross-government program to counter cyberattacks, including those driven by AI, will be headed by Security Minister Dan Jarvis.
Narayan said two government pilots used frontier AI to identify and remediate over a thousand vulnerabilities in public services. A central monitoring service trimmed the time to patch known weaknesses from 50 days to eight.
Ministers will work with Parliament on the Cyber Security and Resilience Bill, including reporting serious AI incidents and near misses. Officials will also formulate methods to manage agentic AI incidents and assist those affected by them.
Security Institute testers now evaluate Google’s Gemini 4 Argon
Over recent weeks, Britain worked with the US government and the labs to keep pre-release access to the most powerful models for the institute, Narayan said.
“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,” he said.
Developers provide models to the institute under voluntary agreements before deployment.
The White House reportedly asked OpenAI and Anthropic last month to hold off on new models for British testers until a US review. The Telegraph reported that Narayan lobbied, and the US changed course.
Ministers have never confirmed that the institute lost access more broadly. “When you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome. 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,” Narayan said.
Cryptopolitan reported in September that Anthropic withheld its Claude Mythos 5.1 model from the institute prior to release. Back in May 2026, the White House was already weighing pre-release reviews of the most powerful models.
Britain will next year hold the presidency of the G20, and AI will be a major focus, Narayan said. He said the government would publish a first look at technical standards Monday afternoon, with the institute playing a prominent part.
The smartest crypto minds already read our newsletter. Want in? Join them.