Britain's AI minister will tell frontier labs the state cannot rely on their self-policing

The speech joins two arguments that answer different problems. Narayan's is about oversight: according to the Guardian, he will argue that Britain cannot leave technology companies wholly in charge of policing their own models, and that the UK needs more AI infrastructure of its own ^[1]. Irving's is about theft. The cyber-defence strengthening that Narayan will announce sits between them ^[2], and the Guardian does not say what the measures cover.

Irving's case rests on one incident. Cybersecurity researchers "hacked" OpenAI under a programme that rewarded ethical hackers, and they said "the scope of what we could theoretically access was huge" ^[17]. The word "theoretically" places the finding at the level of access, which is a weaker claim than a theft. A bounty programme exists to surface exactly this kind of finding. In our view the incident shows what a determined outsider might reach, and says little about what any state has already taken.

Irving is a former OpenAI and Google DeepMind employee who now runs an AI safety research body ^[4]. He has also warned of a "50% chance" of humanity being wiped out by superintelligent AIs, and says the only answer for now is to stop developing advanced systems as soon as possible ^[5]. Someone holding that forecast is not a neutral witness on lab security. But the security claim is narrower than the forecast. It asks whether a hostile state could steal a model, and it can be checked against the controls the labs describe, whatever one makes of the 50% figure.

The decision in front of Narayan is one of sequencing. At least seven countries are named as signatories to the call for coordinated safety standards ^[15], and the speech will be watched for signs the UK may join ^[13]. Last month the prime minister told the UN that Britain would use next year's G20 presidency to push for agreement on AI safety, built on "a single set of global principles and standards" ^[8]. Joining now would give Britain a place in an existing group and commit it before the G20. Waiting keeps its position open for the presidency and leaves it outside the group until then. The reported wording, that Britain cannot rely entirely on companies, stops short of committing the UK to common standards, and we would revise that view if the speech names the coordination call.

Industry is also calling for containment. Satya Nadella is chief executive of Microsoft, which operates AI infrastructure and is an OpenAI shareholder ^[11]. According to the Guardian, he said the starting principle should be to assume that advanced AI models pose an insider risk ^[12]. His point concerns what models do once built, where Irving's concerns who can take them. Nadella wrote: "We must build contained systems whose behaviour we can observe, limits we can test, and actions we can always contain." ^[14]