When the rug gets pulled: why UK safety systems should not depend on US frontier AI

A model here today, gone tomorrow

In June 2026 a directive from the US administration compelled Anthropic to disable its frontier models Fable 5 and Mythos 5 for every user worldwide, only days after they became available. The provider could not restrict access by nationality, so the simplest route to compliance was to switch the models off for everyone. For UK and EU organisations that had started to build on those models, access vanished overnight. There was no comparable domestic frontier model to fall back on.

For anyone designing or installing safety-critical systems in perimeter security and access control, this is the scenario worth taking seriously. If a core capability sits behind a US-hosted API, a single policy change in Washington can remove it without notice. That is not a hypothetical risk now; it has happened once, and the mechanism that caused it has not gone away.

Why there is no easy UK or EU fallback

The obvious response is to switch to a European or British equivalent. The difficulty is that no UK or EU frontier model currently matches the leading US systems for general capability, and the regulatory environment on this side of the Atlantic adds its own friction rather than smoothing the path.

The EU has the more prescriptive regime. The EU AI Act applies a risk-based set of obligations on a phased timeline, with bans on certain unacceptable-risk uses taking effect first and most high-risk requirements following later in the rollout. Proposed simplifications, together with further guidance under consultation, may shift some deadlines. That is helpful in the long run but prolongs the uncertainty for anyone planning a deployment now.

The UK has taken a different route. There is no single AI statute. Instead, existing sector regulators apply principles of safety, transparency and accountability. The Health and Safety Executive has set out its regulatory approach to AI, requiring risk assessments for workplace impacts, controls reduced to as low as reasonably practicable (ALARP), cybersecurity measures, and human oversight under existing health and safety law.

Standards are catching up but are not yet fully in place. CEN-CENELEC's JTC 21 is developing the harmonised standards that will support presumption of conformity with the AI Act, including prEN 18286 on AI quality management systems. In the meantime, BS ISO/IEC 42001:2023, adopted UK-wide through BSI, gives a certifiable AI management system framework that supports both the UK principles-based approach and EU conformity work.

Offline-first: keep the safety-critical path local

We have argued before that a safety system should not depend on the internet, and the Fable 5 withdrawal reinforces the point. An offline-first, edge-native architecture keeps the functions that matter on local hardware. The cloud is still useful, but it is deliberately kept off the safety-critical path.

Dynamic Safety's SAiFI is built on this principle. On-device vision AI handles real-time hazard detection and physical interventions locally, and the system is designed to keep working if connectivity drops. This design directly addresses vendor lock-in, connectivity failures, air-gapped requirements and data sovereignty.

• •Local detection and control loops keep working during internet outages, maintenance windows and vendor changes.
• •Air-gapped deployment becomes possible for sites that cannot, or will not, expose safety functions to the cloud.
• •Keeping data on site supports UK and EU data sovereignty requirements.
• •A cleaner safety case follows from deterministic, local behaviour aligned with ALARP principles.
• •Local event storage with sync-on-reconnect reduces downtime on industrial sites with variable networks.

Self-hosting and on-premise training

Beyond keeping inference local, organisations are looking at self-hosting and training their own models, often using open-weight systems on local edge hardware. This is a direct hedge against sudden access revocation: a model you host and own cannot be switched off by someone else's directive. It also gives you full control over data governance and conformity work.

The trade-offs are real, however. Self-hosting carries higher compute and expertise costs, site-specific training overhead, and ongoing model maintenance. Smaller edge models can also fall short of the peak accuracy of the largest frontier APIs on general tasks, though for a narrow, well-defined safety function (such as detecting a specific hazard interaction) a custom local model can perform very well.

Weighing the alternatives

Industry guidance points the same way. The Security Industry Association makes the case for edge AI in physical security, including low-latency local processing and interoperability. Resilience testing, graceful degradation, and keeping core functions local are not optional extras for safety-critical work; they are the difference between a controlled response and a silent failure when something upstream changes.