Home

Donate
Perspective

Britain’s AI Assurance Approach Needs a Second Act

Luke Cavanaugh, Harry Camilleri / Nov 17, 2026

The authors are writing in their personal capacity.

Shutterstock

Republish

Since its establishment in November 2023, the United Kingdom’s AI Security Institute (AISI) has taken on a growing role in the UK government’s efforts to assess the risks posed by advanced AI systems. With headlines increasingly dominated by AI agents hacking governments, terrorists attempting to weaponize large language models, and calls for pauses in AI development, its establishment looks like a more shrewd bet with each passing day.

Over the past three years, AISI has been highly effective. It has identified dozens of vulnerabilities through pre-deployment evaluations of frontier models, seen its in-house evaluation tools adopted by AI safety non-profits across the globe, and has effectively funneled top technical talent into government (despite offering lower salaries than the private sector). Partly in recognition of this success, AISI-like organizations have been set up in 10 other countries, including France and Japan, often with UK assistance.

But AISI’s current operating model, which revolves around maintaining access to the world’s most advanced models before they are deployed, has been under threat. AISI has no legal power to compel companies to share models, relying on the goodwill of frontier labs, a status quo that would require legislation to change.

This goodwill is no longer certain. Even as Anthropic’s CEO Dario Amodei called on governments to “pace the frontier” in September, Politico reported the White House had asked the frontier labs to withhold new models from AISI until US agencies had completed their own reviews. Although access was restored after 17 days, it is clear that the institute can no longer take privileged access for granted.

This is a cause for concern as AISI seeks to retain its relevance and impact in ensuring the safe development of AI, especially given growing competition from other public bodies and US nonprofits. But the institute has much to offer beyond pre-deployment testing. As it looks to evolve its mission in an increasingly securitized and geopolitically delicate AI landscape, there are three directions in which it could deepen its work.

While any shift in emphasis away from pre-release evaluations would free up capacity, AISI must avoid overstretching its technical staff, its most critical asset. It should focus where it can add unique value, rather than, for example, becoming a research funder or spearheading public sector deployment of AI.

1. Anchor the UK’s domestic AI assurance ecosystem

The global AI assurance industry — from model evaluations to verification technologies — is poised to explode in importance as the technology becomes ever more deeply embedded in national economies and critical infrastructure. Projections for the global market have risen as high as $276B by 2030.

The UK is already well-placed to benefit from this growth. An organic ecosystem has emerged around AISI and Rishi Sunak’s Bletchley Park AI Summit in 2023 — from think tanks like the Centre for the Governance of AI (GovAI) to firms like Apollo Research and Faculty AI — creating 12,500 jobs and adding more than £1 billion a year to the economy.

The government’s 2025 Assurance Roadmap set out an ambitious vision to professionalize the market and boost innovation, calculating that the UK’s domestic assurance industry could be worth nearly £20 billion a year by 2030 with the right interventions. AISI can support these efforts by helping companies become trusted providers, complementing initiatives like an £11 million AI Assurance Innovation Fund to develop innovative technologies. Playing this role would take AISI beyond its current remit as a security research body, and would require a conscious political decision to widen its mandate. But no other public body is currently better placed to serve this function.

First, it can create the infrastructure. AISI’s open source evaluation framework, Inspect, is already used by several other national safety institutes and independent evaluators, including in the US. As the sector grows, the tool could become the established industry standard for public bodies, non-profits, assurance firms and labs. This would require sustained government support, but would significantly lower costs for new firms entering the market, while giving AISI an ongoing role in shaping the field. It would also give British firms, with access to a talent pool trained on the same stack and opportunities to collaborate with the teams refining Inspect, a distinct advantage.

Second, it can help anchor demand for domestic firms by supporting government procurement. Certain use cases, especially in critical national infrastructure, will require sovereign assurance — domestic firms that can meet a high threshold of fidelity in evaluating models. AISI is well placed to help establish the requirements for this type of assurance.

The UK Accreditation Service (UKAS) could then certify firms meeting these requirements, creating a predictable roster of potential suppliers. Certification would also serve as a mark of quality, helping domestic assurance firms win private contracts in addition to their public sector work.

2. Deepen its role as a research hub on key international AI safety topics

Second, AISI should continue to advance the global AI assurance agenda in emerging threat areas. The Institute’s talent base, coupled with the fact that it is not a product of either AI superpower, means it is uniquely placed to play an honest broker role, especially in highly technical domains.

Several critical AI security themes remain underdeveloped, and industry actors have already called for the UK to make a bold investment in AI verification. The same is true of agentic trust infrastructure, the ability to identify, verify and trace agents’ actions. AISI should develop a research agenda for one or more of these “big bets,” in partnership with ARIA, the country’s moonshot R&D agency. AISI would feed into research prioritization and lend its technical teams to support ARIA projects, while ARIA oversees program funding and delivery.

This would be globally useful, but also nationally significant. Any future international agreement on AI safety will need a robust technical underpinning. If AISI can play a role in shaping these discussions, it will not only allow British interests to be represented, but will also ensure the UK’s ongoing relevance as standards develop over time.

More immediately for British businesses, AISI’s research might help develop credible UK assurance standards, reducing the cost and time barriers of regulatory uncertainty for companies looking to demonstrate that their AI systems are safe to deploy.

3. Support critical AI partnerships with other countries

There is much talk, particularly in Europe, about coalitions of middle powers. This conversation is usually framed in terms of groups of nations (the EU, plus the likes of Canada, Australia, or South Korea) banding together to develop pooled leverage or infrastructure.

But the AISI network provides the foundations for a different kind of middle power coalition. Middle powers are the places most directly exposed to AI risk and misuse, lacking their own frontier capabilities but with wealthy populations that make them attractive targets for would-be cybercriminals or terrorists.

These nations must step up cooperation on AI safety . Primarily, this could be achieved through joint defensive acceleration efforts (strengthening society’s defenses against AI risks rather than slowing overall AI development), and these efforts could be delivered through the international AISI network. The UK, for example, could work with South Korea on synthetic media risks or Japan on cybersecurity agents — given the existing focuses of each country’s AISIs.

Not only would this have practical benefits, but it would also represent a form of middle power cooperation palatable to the US. Washington has made clear that it still welcomes working with allies against concrete security threats. Participating countries could even present this as middle powers taking more responsibility for their own security, much as European allies have raised their defense spending.

A program might run in a similar model to the UK Ministry of Defence’s Taskforce RAID, a £100 million body working to integrate AI capabilities into the British military. RAID works on six-to-nine-month mission-focused sprints to build a concrete AI capability or intervention for a tightly defined problem. In this case, each sprint would bring together technical teams from participating AISIs to develop and test a specific defensive capability in a parallel, focused, time-limited setting.

This initiative need not be limited to existing AISIs. The EU AI Act requires every member state to designate a national competent authority for implementation and market surveillance. AISI could help these new bodies build capacity in using tools like Inspect and its wider evaluation methods, embedding UK approaches and standards across different jurisdictions and creating potential markets for British assurance products.

AISI’s capacity is significant but not infinite. The proposals here are a menu rather than a program: AISI is unlikely to pursue them all, but taking up even one or two would mark meaningful progress. Some of these are easier wins than others. Work on Inspect and a joint research agenda with ARIA could start now, as could deeper collaboration through the AISI network. Anchoring the domestic assurance industry would represent a bigger deviation from AISI's existing mandate.

Whatever direction AISI takes, its contribution to the UK and the wider world need not be limited to testing frontier AI models. The challenge will be to decide where AISI can make the greatest contribution and how far it should extend its role.

Support Tech Policy Press
If you've found our work helpful, consider supporting us.

Authors

Luke Cavanaugh
Luke is Head of Strategic Initiatives at Cosine, helping build and deploy sovereign AI models for high-security environments, and is the founder of interweave.gov - a convening platform and think tank for digital government leaders. He previously worked as an Advisor to the Government of Ukraine in ...
Harry Camilleri
Harry Camilleri is an advisor to the Government of Ukraine. Beforehand, he advised the Government of Albania on technology and geopolitics. He has worked on global governance at the think tank E3G, and was a visiting researcher on AI policy at the Leverhulme Centre for the Future of Intelligence.

Topics

Related

Analysis
Burnham's First Week Redraws the Map of UK Tech GovernanceJuly 29, 2026