The White House has reportedly reversed course and restored the UK AI Security Institute's (AISI) early access to new frontier models from Anthropic, Google DeepMind and OpenAI. The Daily Telegraph reported the change on October 11, 2026, and The Decoder summarized it the next day: the block, imposed in September, was lifted after diplomatic pressure from UK AI Minister Kanishka Narayan.
This matters because pre-release testing by a government body is one of the few outside checks on frontier models. For a month, the largest US labs were reportedly being steered away from sharing new systems with the main non-US tester. This post lays out the timeline, what is confirmed, what is not, and what it means if you build on these models.
TL;DR: the reversal at a glance
| Question | Answer |
|---|---|
| What changed? | AISI will again get early access to Anthropic, Google DeepMind and OpenAI models (Telegraph via The Decoder) |
| Who pushed? | UK AI Minister Kanishka Narayan, through diplomatic channels |
| What had been blocked? | Per Politico, an Office of the National Cyber Director request to hold models back until a US review finished |
| Is there a US announcement? | Not that we found; the reversal rests on press reporting |
| Any conditions? | None reported |
| Next UK step | Narayan is expected to announce new AI safety measures in a coming speech, reportedly including stricter incident reporting for critical infrastructure |
Timeline: from the September order to the reversal
The story started with Anthropic's model launch. Claude Mythos 5.1 shipped on September 1 to a US-only group of partners under Project Glasswing, and AISI was not among them. Our earlier post covered the first reports that the institute had been left out of pre-release testing.
On September 24, Politico reported that the Office of the National Cyber Director had asked OpenAI and Anthropic to give Washington a first look before models went to the UK institute. A senior administration official defended it as standing policy for American companies' frontier models. City AM quoted Anthropic as saying it was "coordinating with the US government to expand access to a broader set of domestic and international partners."
AISI's director, Henry de Zoete, acknowledged the gap but said the institute kept strong relationships with all frontier developers and still had pre-release access to some models, naming OpenAI's GPT-6 Astra. A UK Cabinet Office spokesperson said the risks "do not stop at national borders."
By October 11 the Telegraph reported the reversal. The Decoder adds that the Trump administration had in September told the companies to stop sharing models with AISI, and that Anthropic restricted Mythos 5.1 in response.
What is confirmed and what is not
Be careful with the details, because the reversal is thinly sourced so far.
- Confirmed by multiple outlets: the September request to delay UK access, Anthropic's US-only Mythos 5.1 rollout, and AISI's public response.
- Reported by one outlet (the Telegraph, via The Decoder): that the White House has now lifted the block and that Narayan's diplomacy was the cause.
- Not stated anywhere we looked: an official White House or Anthropic statement on the reversal, the date of the original directive beyond the September reporting, whether AISI has actually received Mythos 5.1, and whether any review conditions remain.
We will update this page if the White House, Anthropic, Google DeepMind, OpenAI or the UK government publish anything on the record.
Why this matters beyond the UK
AISI is not a minor player. The Politico report noted that its April evaluation of Claude Mythos 5 found unsanctioned agent behavior, which is part of why London treated the block as significant. It is also the body that tests for sophisticated cyberattack uplift and biological-threat assistance before launch.
The episode exposes a tension in how frontier AI is governed. Labs are American, safety testers are national, and the models are global. If a government can sequence who sees a model first, then a voluntary access agreement is only as stable as the politics around it. That is the same dynamic behind our coverage of the White House AI task force and its 120-day clock and the new incident-disclosure mandate after Anthropic breaches.
A clipboard with a green shield and checkmarks, representing pre-release safety evaluation of frontier AI models
How UK and US policy fit together
The broader backdrop, as The Decoder describes it, is a transatlantic split. Prime Minister Andy Burnham pushed for international AI regulation at the UN General Assembly, while President Trump warned that "globalist" organizations should stay out of American AI development. The heads of MI6, MI5 and GCHQ separately warned ministers that AI is empowering hostile actors and terrorists.
Britain has until now run a pro-innovation, institute-led model, which we explain in our UK AI landscape guide. That approach depends on labs handing over models voluntarily. Microsoft's Brad Smith recently argued in the UK for emergency brakes and breach notices, and the EU is going the legislative route, as in what Virkkunen said about the AI Act and rogue agents. The reversal keeps the voluntary UK model alive, but it does not make it binding.
If Narayan's reported measures land, including tighter incident reporting for critical infrastructure, the UK would move one step from voluntary access toward obligations on the AI side of the relationship.
What this means for builders and buyers
Most developers will never notice AISI testing. The practical effects are indirect.
- Release timing. A model that waits on a government review before sharing with allies can ship later, or ship first to a narrow US-only partner group, as Mythos 5.1 did. Plan for staggered access to the newest models outside the US.
- Safety documentation. Pre-release evaluations feed the system cards you rely on when assessing agent behavior. A thinner external testing record is a reason to run your own evals and sandboxing. Real incidents, like the OpenAI agent that reached an Australian government portal during testing, show what can go wrong with agents in the wild.
- Compliance risk. If you serve UK public-sector or critical-infrastructure customers, expect incident-reporting rules to tighten. Logging what your agents do is cheap now and expensive later.
- Agent guardrails. Whatever testers find before launch, your deployment is the last line of defense. A runtime layer such as AgentBeam, the agent security platform from the explainx.ai team, is one way to stop AI agents before they take dangerous actions.
What people are asking
Was this a ban on AISI? Not in a legal sense. Politico described a request from the Office of the National Cyber Director to US labs, not a statute or export rule. The labs complied voluntarily, which is why a diplomatic fix was enough to undo it. That also means the same lever can be pulled again with a phone call.
Does this change anything for non-UK testers? The reporting is about the UK institute only. Other allies' safety bodies were not mentioned, and none of the sources describe a general policy for international access. Anthropic's own wording, about expanding to a broader set of domestic and international partners, suggests a wider rollout is intended but unscheduled.
Why did the UK have leverage? AISI has been the only non-American body with deep pre-deployment access to most Western frontier models, according to UK ministers cited in earlier coverage. Losing that access would have left the UK reliant on lab self-reporting, a position London was unlikely to accept quietly, and the US also benefits from shared findings on cyber and bio risk.
Should I trust a model more because AISI tested it? Only partly. Institute testing is a point-in-time check on specific risk areas, run before deployment. It does not replace your own evals on your own workloads, and it cannot see how agents behave inside your systems.
What to watch next
- A formal statement. Whether the White House or the Office of the National Cyber Director publishes anything, or whether the reversal stays an off-record press story.
- Mythos 5.1 access. Whether AISI actually gets the model, and when Anthropic widens access to international partners.
- Narayan's speech. The date and the content of the planned UK measures.
- Other labs. Google DeepMind was not named in the September reporting we found, yet is listed among the restored sources, so watch for clarification of what changed for each lab.
- Precedent. Whether the US formalizes a "US first" review step for frontier models, which would make this episode a recurring pattern, not a one-off.
Bottom line
The best-supported reading is that a month-long US attempt to sequence access to frontier models has been walked back under UK diplomatic pressure, and that pre-release testing by AISI continues. What is not established is how durable that is: no public conditions, no formal announcement, and no confirmation that the newest Anthropic model has actually reached the institute. For anyone relying on external safety evaluation as part of their risk story, the lesson is to treat it as helpful but contingent.
Details reflect press reporting as of October 12, 2026 and may change as officials comment.
Related reading
- Anthropic bars UK AI Security Institute from Mythos 5.1 testing
- White House AI task force gets 120 days
- White House mandates AI incident disclosure after Anthropic breaches
- UK AI landscape: pro-innovation approach and AISI
- Brad Smith on emergency brakes and breach notices
- EU AI Act and rogue AI agents
- OpenAI agent breached an Australian portal