It’s been a busy month for MPs racing to catch up. Jacob Coxon made global headlines warning that we could all die within the decade. Rogue OpenAI agents infiltrated Australian government systems. False intelligence in an AI-generated report almost prompted the US military to attack a Chinese ship in the Middle East. In Parliament, MPs debated a new Superintelligence Bill, while leaders from over 20 countries signed a call to keep frontier models under human control.
Last week, the government set out a package of measures in response. The AI Minister, Kanishka Narayan, said he wants Britain to be "one of the hardest countries in the world to attack using artificial intelligence". He had a blunter message for the frontier labs: if you cannot control the risks a model poses, do not release it.
Both are welcome ambitions. But delivering them will require Britain to confront an uncomfortable reality: we have limited power over the companies developing the most capable AI. This is important because the most significant security breaches to date - the Australian government hack, and the now infamous Hugging Face scandal - were carried out by models before they were deployed. AISI’s pre-deployment access to evaluate the most advanced models has historically been Britain’s strongest lever to influence model development. But events in September revealed this to be a precarious arrangement: the White House restricted AISI’s access, only to restore it after lobbying from British ministers. We are one White House decision from being cut out of the loop.
And other steps MPs are considering - like the Superintelligence Bill or calls for an international treaty to slow development - are difficult for Britain to enforce, as companies in the US or China have few incentives to consider British interests when developing their models. Britain risks becoming Cnut commanding the tide: asserting authority over forces we have little power to shape.
But dependence need not mean powerlessness. The risks will arrive on our shores whether we capture value from AI or not. Rather than retreating from advanced AI, the more useful approach is to question what tools Britain has to reduce our exposure to AI risk. Thankfully, there are high leverage bets well within our power.
Making AI safety a diplomatic priority
Britain should start by building a diplomatic agenda around safer AI. We have well-established convening power on AI safety since Bletchley in 2023. Andy Burnham’s recent proposal for Britain to act as an ‘honest broker’ in developing global AI standards is welcome. He should establish a Diplomacy Unit within the AI Taskforce, bringing together technical experts from AISI and Foreign Office officials to set the strategic direction and draft standards for safe frontier development.
The key mechanism to deliver an AI diplomacy platform will be Britain’s G20 presidency.
There are two avenues to pursue. First, Britain should bring the US and China into discussions on shared safety standards, including red lines around AI systems autonomously improving their own capabilities (known as recursive self-improvement). Second, building on the US G20 presidency’s focus on secure supply chains, Britain should establish a G20 working group on the AI supply chain. It should invite the Netherlands as a guest and work closely with Japan and South Korea. These countries make the lithography machines and memory chips that frontier AI depends on, which gives the group leverage that Britain lacks on its own. The group should push for access to frontier AI, including for AISI’s pre-deployment testing. Access is increasingly a security matter, as nations come to rely on frontier technology to defend themselves. As Australia discovered, autonomous agents can now breach government systems before leaving the AI companies, so even well-meaning allies can pose security threats to the unprepared.
Leveraging British expertise to make agreements verifiable
Compute verification technology is another fruitful area for Britain to leverage its technical expertise. If any slowdown agreement is to be reached between the US and China, it will depend on each side verifying that the other is upholding the terms of the agreement, and not secretly undertaking large-scale training runs. The technology to verify that is a nascent field, and the UK has the opportunity to lead it. The government agrees, and has promised details of support for UK verification efforts in the coming months. That support should start with scenario planning (as Ed Lawrence and I have previously outlined) to determine exactly what verification hardware will be needed, then use advance market commitments to de-risk the efforts of hardware innovation companies such as Sheffield-based Amodo. Britain could lead a group of allied nations to pool technical expertise and ultimately help facilitate a slowdown agreement; the US and China are unlikely to trust technology developed solely by the other, so this is a promising area for nations like Britain to make an outsized contribution.
Giving the state tools to act in a crisis
Domestically, Monday's package is a good start. The government will explore using the Cyber Security and Resilience Bill to protect hospitals and utilities from AI risk, including loss of control, and look at how serious incidents and near misses are reported. While this protects critical services exposed to AI risk, it leaves us with few powers over the companies creating it. Consider the three weeks when AISI was cut off: when Washington held back advanced models, even Brussels had more tools than the UK to influence the frontier AI companies, as they could at least legally demand information from them.
MPs should close that gap by amending the Cyber Bill to give ministers emergency powers to demand information from frontier AI companies and restrict the deployment of dangerous models. Without them, the government’s best option in a crisis is a 2004 emergency law, the Civil Contingencies Act, which has never been tested and is open to legal challenge. Incident reporting should also cover AI companies, not just operators of critical infrastructure. The companies’ staff and assets in Britain incentivise them to cooperate, even if British legislation has little influence over model development overseas. Scoped to national security emergencies, these powers would not slow growth. Guardrails like these are what make aggressive adoption defensible.
Preparing critical infrastructure for new threats
Finally, the AI Taskforce should lead a preparedness programme across critical infrastructure and government departments. This is an agenda that the Cabinet Secretary Antonia Romeo is well-placed to coordinate with technical staff at her disposal. She could start by putting loss of control on the National Risk Register, so every department makes contingency plans for it. The government's pilots using frontier AI to scan its own systems have already found and fixed over a thousand potential vulnerabilities, and are now being extended. That approach should be mirrored by critical infrastructure operators too. The Taskforce should set standards for defensive cyber capabilities and ensure that critical infrastructure can switch between model providers, so defences hold when access to a specific model is lost.
Britain should also invest in domestic open source capability, and enough domestic inference compute to run it, to act as a defensive backstop when access to American technology is cut off. This may not require the most advanced models: some specialised providers already claim to rival frontier systems at finding and patching vulnerabilities. The Defence Investment Plan earmarks £115m to protect against AI-enabled threats; the Taskforce should lead how that money is spent. Biosecurity should be a core part of this work. One cheap win is expanding biological monitoring to enable early warning of emerging threats.
Don’t mistake ambition for leverage
Monday's package was a welcome first step. But the risks are advancing faster than Whitehall can plan for them, and Britain is not yet prepared for the new normal.
The real Cnut sat before the tide to show his courtiers the limits of his power. Britain is on course to make the same demonstration. Avoiding it means using every tool we have, leveraging our weight on the global stage and strengthening resilience at home.