The UK's AI safety testing was switched off for a fortnight — and that is the real story

The UK’s AI Security Institute will once again get pre-release access to the most capable models from OpenAI and Anthropic, AI Minister Kanishka Narayan confirmed on Monday, after Washington lifted a freeze that had cut British safety testers off from the newest systems for around two and a half weeks. The access was reinstated on 11 October, according to reports; the pause had begun in late September, when the US administration told the labs to hold their latest models back from the institute while American authorities completed a review.

It is tempting to file this as good news and move on — the block is gone, the testing resumes. But the more important story is the one Narayan himself told, with notable candour: for a fortnight, the institution the UK holds up as the centrepiece of its AI safety credibility simply could not do its core job, and there was very little Britain could do about it.

What the AI Security Institute is for

The AI Security Institute — the renamed successor to the body the UK launched with great fanfare in 2023 — exists to test frontier models before they reach the public, probing them for dangerous capabilities and failure modes. We have covered its work closely, from red-teaming safety monitors to stress-testing AI agents for unsanctioned behaviour. All of that depends on one thing: getting access to the models early, before release. Take the pre-release access away and the institute is left evaluating systems the public already has, which is a fundamentally weaker position.

That is exactly what happened. When the models were withheld, the institute’s distinctive advantage — seeing frontier systems first — evaporated overnight, through no decision of its own.

The sovereignty problem in plain sight

Narayan’s own words were the most revealing part of the announcement. He said he was confident the institute “will be able to test the latest generation of models,” but added the crucial caveat: “When you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome.” That is a minister acknowledging, on the record, that a core pillar of UK AI policy runs on access that another government can switch off.

This is the uncomfortable thread running through so much of the UK’s AI position, and it connects directly to the country’s sovereign-AI push. Britain has world-class evaluation expertise, but almost all the frontier models that matter are built by American companies, subject to American policy. The UK’s own frontier lab, Google DeepMind, is headquartered in London but owned by a US parent. When safety testing, compute, and the models themselves all ultimately depend on decisions made in Washington and San Francisco, “clear-eyed about permanence” is about as honest as a minister can be.

The rest of the speech: control and cyber

Speaking at Imperial College London, Narayan paired the announcement with a firmer line on the labs themselves. His message to frontier developers was blunt: “If you cannot adequately control the risks, do not release the system.” It is a stance consistent with the UK’s broader principles-based approach to regulation — no single AI law, but a clear expectation that those building the most powerful systems carry responsibility for controlling them, especially as models begin to act with less human oversight as AI agents.

He also announced new cyber-defence measures, including government pilots that used AI to identify more than 1,000 potential vulnerabilities in the UK’s own systems and to speed up fixing them — a rare example of the state using the technology defensively on itself rather than only regulating its use by others.

Why it matters

The freeze was short and it is over, but it was a useful stress test of a different kind — of the UK’s assumptions. It showed that the AI Security Institute, for all its expertise, operates at the discretion of foreign companies and a foreign government, and that the UK’s safety regime is only as sovereign as its access to the models it is meant to police. Narayan’s candour is welcome; the dependency it describes is not going away. The question the episode leaves hanging is what, concretely, the UK intends to do so that the next freeze — if it comes — is less of a problem.