AISI Regains Access to Test Anthropic, DeepMind and OpenAI Models Before Launch
According to more recent reporting, the White House has withdrawn the September decision that shut the UK's AI Security Institute out of early access to models from Anthropic, Google DeepMind and OpenAI.
The reversal comes after diplomatic pressure from UK AI minister Kanishka Narayan – just before he is due to deliver his first major speech on AI safety since the portfolio was elevated to government level.
What Has Changed
According to The Decoder, which cites The Daily Telegraph, the White House has lifted the September block, when the Trump administration reportedly asked the American AI labs to stop sharing models with the UK's AI Security Institute (AISI). The institute tests AI systems for safety risks before they launch, and is now reportedly to regain early access to models from Anthropic, Google DeepMind and OpenAI (The Decoder).
It is worth emphasizing what the reporting actually rests on: The Decoder cites The Daily Telegraph, and no primary documents – statements from the White House, the UK government or the labs – appear in the source material. It is also unclear whether all three labs have already restored access, or whether they have only pledged to do so.
The September Block and Claude Mythos 5.1
Back in September, the Trump administration reportedly instructed the companies to stop sharing models with AISI. According to the same report, Anthropic subsequently restricted access to its new model Claude Mythos 5.1 (The Decoder). For AISI, this was a concrete narrowing of its working basis: the institute's function is precisely to test frontier models for safety risks before they reach the market, and without early access, testing falls away at the most vulnerable moment – before launch.
That an allied state was shut out of safety testing of American models points, according to the reporting, to the tension between UK AI oversight and the US administration's national approach to AI development.
The Diplomacy Behind the Reversal
The Decoder writes that the block was lifted after diplomatic pressure from AI minister Kanishka Narayan. The report places the matter in a broader diplomatic context: Prime Minister Andy Burnham had earlier pushed for international AI regulation at the UN General Assembly, where Trump warned that "globalist" organizations should stay out of American AI development. At the same time, the heads of MI6, MI5 and GCHQ had warned government members that AI gives hostile actors and terrorists new opportunities (The Decoder).
It is uncertain exactly how the pressure was applied, and the White House's motivation beyond the reported diplomacy is not documented.
Narayan's Upcoming Measures
The reversal comes just before Narayan, according to several sources, is due to deliver a major speech on AI safety – his first since AI became a government-level portfolio under Prime Minister Andy Burnham (HT World). Key elements of the upcoming measures:
- Stricter reporting requirements for AI-related safety incidents involving critical infrastructure (The Decoder).
- A call to frontier developers to do more to secure their systems, and an argument that the UK cannot fully and entirely trust technology companies to oversee their own systems (HT World).
- More British AI infrastructure, so the country is not entirely dependent on foreign actors to evaluate and defend against AI risks (HT World).
The Times has previewed the speech under the headline "'Inevitable' that Britain will be hacked by rogue AI, says minister" – according to the subheading, Narayan is to ask technology companies not to release systems they cannot control, while being described as the one who has drafted Britain's first safety laws against the threat (The Times). The content of these laws has not yet been published in detail beyond this.
Why AISI Access Matters
AISI represents a model several countries have looked to: a state body that gets to see frontier models before launch and tests them for dangerous capabilities. For the labs, this provides external validation; for the British authorities, it provides early insight into risks before they materialize in products used in UK societal infrastructure.
The September block broke with this model. The restoration, if it holds, means the UK retains its place in the early testing work – but the dependency relationship is unchanged: the institute remains dependent on the goodwill of the American authorities and the companies' willingness to cooperate. That is likely the backdrop to Narayan's argument for domestic AI infrastructure.
Counterpoints and Open Questions
Several factors mean the story currently rests on uncertain ground:
- No primary sources. The reversal has been reported by The Decoder citing The Daily Telegraph. Neither the White House nor the British authorities have confirmed the decision in the available documents, and details on timing and mechanism are lacking.
- Restored, or only pledged? It is unclear whether all three labs have actually reopened access, or whether this is an intention. Anthropic restricted access to Claude Mythos 5.1 in September; there is no confirmation that this model in particular is now available to AISI.
- How durable is the reversal? The block came from the White House and was lifted after political pressure. The same mechanism could, in principle, be applied again, leaving the institute in a vulnerable position.
- The content of the UK measures. Both the reporting requirements and the "first safety laws" are currently known only in outline. It remains to be seen how strict they actually are, and how they will be enforced against companies that are primarily under American jurisdiction.
What Comes Next
The next milestone is Narayan's speech, where the restoration will likely serve as diplomatic cover: it shows that British pressure can work, while the speech's message – that the UK must reduce its dependence on both American politics and the companies' self-policing – precisely acknowledges how fragile that dependence is. Until confirmations come from the labs themselves, the reversal should be read as reported, not as documented fact.