A standard-setter that had to ask for access
On 12 October 2026, AI minister Kanishka Narayan told an audience at Imperial College London that the AI Security Institute, or AISI, is “now conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon”. According to the speech text published by gov.uk, that access came “over the last few weeks” of work with the US government and the labs “to retain access for Britain’s AI Security Institute to the most powerful AI models before they are released”. Later in the same speech he said he “will not accept a future in which Britain depends on technologies it cannot understand or control”. The two statements sit in the same speech, and the distance between them is the subject of this analysis. We covered the speech itself as news; here we ask what it means for the UK’s claim to lead on AI safety standards.
Strategic Insight: Britain’s standards ambition rests on AISI’s technical record, and that record rests on pre-release access to American models. Reports in September said Washington had asked the labs to withhold their newest models; reports in October said it had reversed that request. Whatever the UK publishes as a standard, the evidence behind it currently arrives at another government’s discretion.
Four signals in one week
The speech landed in a week when four separate voices described the same dependency from different sides.
The first was Washington’s own reversal. LBC reported on 12 October, citing The Telegraph, that the White House had reversed its request that Anthropic, OpenAI and other labs withhold their newest models from AISI, and that the change followed lobbying by Narayan. City AM had reported the original request in September. It came from Washington’s Office of the National Cyber Director, and a senior US official explained it with the words “because they’re American companies and this has been our policy with every new frontier model that comes out”. Anthropic withheld Claude Mythos 5.1 from the institute, saying that access had at first been limited to US organisations. We reported the request and the reversal at the time.
The second was Microsoft president Brad Smith telling the BBC that the UK is best placed to drive global AI rules because of AISI, and that in Washington “people don’t look at the UK and they don’t look at London the way they look at Brussels”. Smith wants emergency brakes, legally required incident reports and independent government evaluators who vet new AI before release. Our report on the interview has the detail.
The third was George Williamson, who now runs the Alan Turing Institute, telling The Guardian that the UK “shouldn’t depend solely on systems we can’t inspect or control, or that could be restricted or even switched off”. He said recent events showed “both political and commercial pressures mean we may not have access [to AI] in all the circumstances when we want it”. We covered his interview on Monday.
The fourth was The Guardian’s preview of Narayan’s speech, which said he would argue that Britain “cannot rely entirely on technology companies to police their own models”. The paper noted that an international call for common AI safety standards has been signed by the leaders of countries including Ireland, the UAE, Singapore, Australia, Germany, France and Finland. Neither the UK nor the US has signed, and the paper records that President Trump wants the leading labs to “self-police”.
What the week actually revealed
Put together, the four signals describe a country that wants to write the rules, has the technical institution to do so, and depends on a foreign government’s goodwill for the raw material. Smith’s endorsement is real, but it is the view of a company whose own president’s administration reportedly asked the labs to hold models back. Williamson’s warning is about procurement and resilience, but it applies with equal force to evaluation: an institute that cannot see a model cannot set a standard for it.
| Figure | Value | What it tells you |
|---|---|---|
| Models AISI has evaluated before release | more than 55 (AISI blog post) | The technical record Smith is pointing at is substantial |
| Countries The Guardian names among the signatories to the international standards call | seven, with the UK and US absent | The UK has not joined the call its speech was being watched for |
| Vulnerabilities found by two government pilots using frontier AI | over 1,000 (gov.uk) | The cyber-defence half of the speech has delivered results already |
| Time to fix identified weaknesses, per the speech | from 50 days to eight | The same point, measured in speed |
| AISI technical staff, per the speech | over a hundred | A small institution for a job the minister describes as foundational |
Strategic Reality: AISI’s own blog post, the one the speech promised for that afternoon, says it “is not seeking to mandate requirements for frontier developers”. The standards Narayan says AISI will help shape are, on the institute’s own account, voluntary best practice that “could be formalised via international standards development organisations” if practices mature.
What the UK can set, and what it cannot
The voluntary track
The speech promised that AISI “will take a significant role in shaping our collective understanding of technical standards”, and that an initial view would follow that afternoon. It did. AISI’s post lists eight priorities, from documenting risk through the development lifecycle to defining unsafe agent behaviour, enabling independent auditors, and sharing metrics on how fast AI is developing. The post describes a route in which developers and technical bodies iterate on practices first, and formal standards come later. It also welcomes the labs’ commitments under what it calls “the White House Accord on Super Intelligence”.
That is a credible technical agenda, and the post’s diagnosis is blunt: risks now arise “earlier in the development process”, and evaluation methods for alignment and control “can’t assure whether increasingly autonomous AI agents will behave as intended”. But the agenda carries no obligation. Smith’s argument to the BBC was that a self-regulatory regime has a built-in flaw, because “the good companies will do what you want” and “it’s the companies that are not as responsible that you need to worry about”. AISI’s framework, as published, is addressed to the companies that choose to engage.
Critical Context: The Guardian preview framed the speech as a warning that Britain “cannot rely entirely on” firms to police their own models. The words in the speech text gov.uk published were an instruction without a stated consequence: “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 press release’s version of the line differs by a few words. He then added: “I want the labs to be honest.”
The legislative track
Narayan said the government “will legislate wherever necessary”. The one piece of legislation he named is the Cyber Security and Resilience Bill, which the speech describes as an opportunity to strengthen the regime for “AI-driven cyber threats and loss-of-control risks”, including “identifying and reporting on serious incidents and near misses”. Parliament’s Bills API shows the Cyber Security and Resilience (Network and Information Systems) Bill at Report stage in the House of Lords as of 12 October. That is a bill about the resilience of essential services and incident reporting. It is not a bill that gates the release of frontier models in the UK.
The release-gating idea has been on the table before. The Guardian reported in September that ministers under Keir Starmer had explored whether the most advanced companies could be forced “to submit their products for safety testing before launching them”, and that the plans “fell by the wayside”. Chi Onwurah, chair of the Commons science and technology committee, put the sharper version to the same paper: “if your frontier model has not gone through the safety institute, we are not going to allow you to sell your commercial models in our country.” Nothing in the 12 October speech or press release revives either proposal. Our earlier analysis of AISI’s enforcement gap covers that history.
Success factors the speech leaned on
- Access. The speech is explicit that pre-release evaluation of the latest models follows weeks of work with the US government and the labs to retain access. AISI’s blog says it has “independently evaluated more than 55 models before release”, and presents that record as the basis for its expertise. Remove the access and the expertise stops accumulating.
- A coalition that excludes nobody important. Narayan said capable models “distributed openly” make it “impossible to be an isolated island”, and that AI will be central to the UK’s G20 presidency. The Guardian’s September reporting quoted Labour MP Liam Byrne’s view that “a safety regime which is set and enforced by the UK, Europe, India, Canada and Japan would be robust”, and an unnamed former official saying that if enough countries required models to pass through their safety institutes “we would be able to stand up to the US government”. The UK has not signed the existing call.
- Verification that does not need permission. The speech named “hardware-enabled verification, drawing on Britain’s expertise in semiconductor design” and called it “early research”. The press release promised more on AI verification “in the coming months”. This is a strand that could reduce the dependence on voluntary disclosure, and the speech itself calls it early research.
- Domestic capability. Narayan said Britain “will not accept dependence as the price of” inward investment, and pointed to the Sovereign AI programme and the AI Hardware Plan. Williamson, for his part, does not expect the Turing Institute and the UK to match the big labs’ resources, The Guardian reports; he said they could still produce “nationally owned IP” or smaller models for startups.
Why the dependence does not resolve
Williamson’s answer to dependence is to own some systems outright. He describes the aim as “nationally owned IP” or smaller models, “without trying to copy what frontier labs are doing”. The speech’s answer is a mix of negotiation, verification research and domestic building. The speech describes the access as something retained through weeks of work with the US government and the labs, not as an entitlement. The September episode shows the US government deciding what AISI sees, and the City AM report records the US position that this has been “our policy with every new frontier model”. LBC, citing The Telegraph, attributes October’s reversal to lobbying by Narayan; the speech credits weeks of work with the US government and the labs. Either way it is a diplomatic result, not a structural one.
⚠️ Warning: LBC’s account of the reversal does not say when access resumes, whether it covers every lab, or whether it comes with conditions. The speech confirms evaluations of Gemini 4 Argon. Neither source confirms whether Claude Mythos 5.1 has gone to AISI.
Who this reaches
The groups on each side of the access line
Frontier labs are being asked for honesty and restraint, and the levers the speech names are voluntary practice, a cyber bill and a promise to legislate where necessary. OpenAI’s European policy head, Tom Duff Gordon, told The Guardian in September that the company supports “stronger UK rules for the handful of companies” building the most powerful systems, so the demand for rules is not coming solely from government.
UK organisations building on American models sit where Williamson’s warning bites. At the Labour conference, POLITICO reported, Jonathan Reynolds, the business secretary, said the UK must keep its access to the leading models developed outside the country. The September reports show that even the government’s evaluator went without one model before its release. A firm whose product depends on a model it cannot inspect, and that could be restricted, is in the position Williamson described as “not a place where a nation wants to be”.
Hospitals, utilities and the other operators of essential services are the group the legislative track actually touches. The press release says ministers will look at strengthening the Cyber Security and Resilience Bill regime so that hospitals, utilities and other such services “are better protected from emerging AI related risk”, including loss-of-control risk, and will consider how near misses are identified and reported.
| Group | What changed on 12 October | What is still open |
|---|---|---|
| Frontier AI labs | A public instruction not to release systems whose risks cannot be controlled; AISI’s eight voluntary priorities | Whether any UK obligation follows if a lab ignores the instruction |
| AISI | Confirmed pre-release evaluation of the latest model generation, after negotiation | Whether access is now standing, conditional or model by model |
| UK firms built on US models | A ministerial statement that dependence is a national security question | Any mechanism that protects a UK customer’s access when a lab or Washington restricts it |
| Essential-service operators | AI incident and near-miss reporting to be examined under the cyber resilience bill | The detail, which the press release says will be settled with parliamentarians and operators |
What this means for firms building on frontier models
The sources support one practical implication. Williamson’s test, whether a system “could be restricted or even switched off”, sits beside Narayan’s refusal to accept dependence on technologies Britain “cannot understand or control”, and the September reports of a model withheld from AISI show the risk is not hypothetical. Reynolds’ insistence on retaining access to foreign models and Narayan’s refusal to accept dependence are two halves of the same government position. UK firms sit between them, and neither the speech nor the press release announces a measure on the terms on which a UK customer gets or loses access to a US model.
The open challenges
Standards without a lever
AISI’s priorities are voluntary by design, and Smith’s point that voluntary regimes bind only the willing was made on the same day. Narayan’s challenge to critics at the Labour conference, reported by POLITICO, was to show “how your mechanism proposed gets you to the outcome”. The speech does not answer his own question for the release of frontier models. He wrote in the Times, as The Guardian reported in September, that “nothing is off the table”. The table, on 12 October, held a cyber bill and a research programme.
Access that is granted, not owned
The US official’s “our policy with every new frontier model” is the sentence that matters. It describes a standing position, not a one-off. The October reversal, as LBC reports it, shows the position can bend to lobbying; nothing in that report says the position has changed. Narayan’s own words acknowledge this: much of the technology shaping the UK’s economy and security “is being developed outside this country”.
Capability that stops short of the frontier
Both Williamson and the speech answer dependence with domestic building. Williamson, per The Guardian, does not expect to match the big labs’ resources. We examined the money behind Sovereign AI in our analysis of the funding gap. Smaller, nationally owned models are his route to resilience. They do not give the UK a frontier model to evaluate on its own terms.
A coalition that has started without Britain
The Guardian’s list of signatories to the call for coordinated standards does not include the UK. Narayan’s speech says the challenges “cannot be addressed by a single country alone” and promises Britain “will play its part”. Whether that part is joining the existing call, or building a separate G20 track that can accommodate the US, is unresolved.
Reality Check: By its own account AISI has put more than 55 models through pre-release evaluation and has helped shape the practices the labs now follow. That is a genuine lead, and Smith is not wrong to name it. It is also a lead that was interrupted for at least one model in September, and the remedy the reports describe was lobbying.
What to watch
Whether the UK signs. The Guardian preview said the speech would be watched for signs the UK might join the call. The speech text announces no such step. A signature before the G20 presidency would show the UK choosing the coalition over its US alignment; continued absence would show the reverse.
Whether AISI’s priorities move towards formal standards. The blog post’s own route runs through international standards development organisations once practices mature. A first draft submitted to such a body, or a published verification method the labs are asked to adopt, would mark the shift from best practice to standard. The press release promised detail on AI verification “in the coming months”.
The Cyber Security and Resilience Bill in the Lords. The bill is at Report stage. Government amendments on AI incident or near-miss reporting, or on loss-of-control risk, would show the legislative track doing what the speech said it would. Their absence would leave the “legislate wherever necessary” line untested.
Whether the next withheld model stays withheld. LBC’s report does not say whether the reversal covers every lab. The next frontier release will show whether AISI’s access is standing or negotiated again. If Claude Mythos 5.1 is confirmed as evaluated by AISI, that closes the September episode; if the next model is held back, the October reversal was a pause.
The reason this matters now is that the G20 presidency is next year, and Narayan has told the labs to hold back systems they cannot control. A government that says that to companies while negotiating for the right to see their models has a credibility problem it cannot lobby its way out of. Either the UK gets a standing basis for evaluation, through a coalition, through verification methods that do not rely on disclosure, or through a legal requirement for market access, or its standards will be advisory notes written about systems it sees when permitted.
Strategic Insight: The question for the next year is not whether Britain has the expertise to write AI safety standards. AISI’s record settles that. It is whether the UK will attach a consequence to them, and whether the evidence they rest on will be available without asking.
Source: UK to strengthen cyber defences as minister calls for AI firms to increase safety efforts (The Guardian, 12 October 2026), by Dan Milmo and Robert Booth. Additional sources: Building a Safer, More Secure Britain in the age of AI (gov.uk speech text, 12 October 2026); Britain steps up national defences against growing AI risks (gov.uk, 12 October 2026); Towards technical best practice for frontier AI development (AI Security Institute, 2026); White House reinstates UK access to new AI models for safety tests (LBC, 12 October 2026); White House tells AI giants to hold models back from UK safety watchdog (City AM, September 2026); AI is less regulated than milk, says Microsoft president (BBC News, 12 October 2026); UK must not become dependent on foreign AI, says head of Alan Turing Institute (The Guardian, 10 October 2026); ‘A critical moment’: concern UK is not up to speed in acting on AI risks (The Guardian, 18 September 2026); UK AI minister: focus on building tech, not regulating it (POLITICO Europe); Cyber Security and Resilience (Network and Information Systems) Bill (UK Parliament Bills API, status as of 12 October 2026).
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether AISI’s voluntary priorities acquire a consequence, and whether the next frontier model reaches the institute without a negotiation first. Read more analysis at Insights, or get in touch.