The UK government is assessing the economic and security risks of losing access to advanced US artificial intelligence models, after a temporary US export control episode disrupted access to two leading models in June. The review, first reported by the Financial Times, follows a Trump administration intervention that restricted foreign access to Anthropic's Fable 5 and Mythos 5 models, prompting Anthropic to suspend access for all users because it could not verify nationality in real time. The restrictions were lifted on June 30.
The episode has raised a practical question for government departments accelerating AI adoption across public services: what happens when organisations become dependent on models whose availability they do not control?
"The central lesson is that access to an AI model should never be mistaken for control over it," said Chris Newton-Smith, CEO of IO. "This is particularly important as organisations move quickly to adopt AI, often without fully considering the long-term implications."
Fast adoption creates new risks
IO's State of Information Security Report 2025, based on research among more than 3,000 information security professionals in the UK and US, found that 54 percent said their organisations had adopted AI technology too quickly and were now facing challenges in scaling it back or implementing it more responsibly.
Newton-Smith said organisations have often approached AI as though it were conventional software - selecting a supplier, integrating its service and assuming it will remain available on broadly the same terms. But advanced AI models increasingly sit between national security, export controls, commercial policy and geopolitics, meaning availability can change for reasons outside the customer's control.
For public sector organisations, that makes concentration risk and exit planning increasingly important. "Before embedding a model in a critical service, departments should understand its dependencies, data flows, alternatives and how the service would continue if access were restricted," Newton-Smith said. "This does not mean avoiding frontier AI. It means designing for substitution from the outset, with portable data, flexible architecture, contractual protections, tested fallbacks and human-operated alternatives."
Treat AI providers like critical suppliers
Newton-Smith argues AI providers should increasingly be treated in a similar way to critical cloud suppliers and other strategic technology partners, although AI introduces additional considerations. Alongside resilience, security, data handling and continuity, departments need to consider providers' ability to change model behaviour, policies or limitations without any corresponding change to the customer's application.
Assessments should also consider jurisdiction, intellectual property, transparency and security, as well as model-change notifications and whether suppliers can support investigations or redress. Departments need to look further down the supply chain too, as an AI provider may itself depend on other models, cloud platforms or data sources.
The level of assurance should depend on how AI is being used. Newton-Smith contrasts a drafting assistant used for low-risk internal communications with a system influencing eligibility, enforcement or access to public services. "Supplier assurance should be proportionate to the potential impact and criticality of the use case," he said.
As countries take different approaches to regulating AI and controlling access to advanced models, Newton-Smith said organisations shouldn't assume that a single global AI deployment will remain legally and operationally consistent everywhere. Instead, they need visibility of where systems, users and data are located, which suppliers are involved and which models underpin individual services.
For those working in government AI roles, the gap between adoption speed and governance maturity is now a central concern. AI for Government training can help teams build the practical skills needed to assess supplier risk and design for substitution before systems go live.
Building resilience into public sector AI
For government, Newton-Smith said the starting point should be an accurate record of each material AI system, including its purpose, data, dependencies, affected people and accountable owner. Higher-impact systems should be subject to stronger evaluation, human oversight and routes for appeal and redress, while monitoring should continue after deployment to cover model updates, performance drift, security incidents and regulatory change.
Critical public services also need tested continuity plans, alternative suppliers or models, data portability, exit support and non-AI fallbacks. "Above all, AI governance must bring together policy, risk, procurement, data, security, legal and operational teams - not operate as an isolated technology programme," he said.
Why this matters for government professionals
For civil servants and public sector technology leads, the practical takeaway is that AI procurement decisions now carry geopolitical risk that standard software contracts did not. Before committing to a frontier model for a critical service, departments should document dependencies, negotiate model-change notification rights, and maintain a tested fallback that can operate if access is restricted. For policy teams shaping AI strategy, the AI Learning Path for Policy Makers covers the governance and risk frameworks needed to embed these requirements into departmental practice from the outset.
Your membership also unlocks:
