Operators of essential services often treat an AI project as something outside their cyber compliance, a technology initiative to be governed later. In a regulated essential service the model is already inside the security regime the moment it touches an operational system, and the framework that assesses it has already been updated to look for it.
The short answer
- If you run an essential service, the Network and Information Systems Regulations 2018 already apply. An AI system that touches that service falls inside that scope rather than beside it.
- The practical benchmark is the NCSC Cyber Assessment Framework. Version 4.0, released in August 2025, adds explicit coverage of AI-related cyber risk, so your competent authority now expects to see it addressed.
- AI mainly changes the risk picture through new data dependencies, a model supply chain, and attack routes such as data poisoning and prompt injection, all of which map onto existing CAF objectives.
- The Cyber Security and Resilience Bill, before the House of Lords through 2026, will widen scope and put the CAF on a statutory footing. It sharpens the direction rather than reversing it.
- The gap is usually ownership rather than tooling: whether a named person is accountable for each AI component as an in-scope asset.
The regime you are already in
The NIS Regulations 2018 place security and incident-reporting duties on operators of essential services in energy, transport, health, drinking water and digital infrastructure, enforced by sector competent authorities such as Ofgem, with the NCSC as the technical authority. The duty is outcome-based: take appropriate and proportionate measures to manage risks to the network and information systems on which the essential service relies, and report significant incidents.
That phrasing matters for AI. The duty attaches to the systems the essential service relies on, so it does not distinguish between a conventional SCADA component and a machine-learning model feeding the same control room. If a demand forecast, an anomaly detector or an AI-assisted dispatch tool informs the operation of the essential service, it is part of the estate the regulation already covers. The same logic runs through the delivery of the digital commitments in a RIIO-T3 business plan: the regulator treats the digital capability as part of the regulated service rather than a side project.
The framework has already been pointed at AI
Competent authorities assess NIS compliance against the NCSC Cyber Assessment Framework, which is organised around four objectives: managing security risk, protecting against cyber attack, detecting cyber security events, and minimising the impact of incidents. In August 2025 the NCSC released CAF version 4.0, adding more than a hundred indicators of good practice and strengthening coverage of AI-related cyber risk, secure software development, adversary methods, and monitoring and threat hunting.
The effect is that an assessor now has explicit prompts to ask how AI in your estate is governed and secured. Treating an AI deployment as outside the CAF is no longer tenable, because the framework itself directs attention to it. An operator that cannot show where its models sit against the CAF objectives has a visible gap at its next assessment.
What AI actually adds to the risk
AI does not introduce a new duty. It introduces new failure modes that the existing duty already requires you to manage. Four are worth naming precisely, because each maps onto a CAF objective.
| What AI adds | The risk in operational terms | Where it lands in the CAF |
|---|---|---|
| A model supply chain | Third-party models, training data and APIs become dependencies you did not build and cannot fully inspect. | Managing security risk, including supply chain |
| New data dependencies | The model's behaviour rests on live feeds whose integrity now affects the essential service directly. | Protecting against cyber attack, data security |
| Model-specific attacks | Data poisoning, prompt injection and model extraction are routes that conventional controls were not designed to catch. | Detecting cyber security events |
| Automated action | A model that acts, rather than advises, removes the human step that resilience and safety cases often assume. | Minimising the impact of incidents, including manual fallback |
Read down that last column and the point becomes clear. Every AI-specific risk has a home in the framework you are already assessed against. The work is to trace each model through the four objectives and evidence the controls, not to stand up a separate AI security programme beside the one you have.
What the incoming Bill changes
The Cyber Security and Resilience Bill was introduced in November 2025 and, through 2026, has been progressing through Parliament, reaching the House of Lords in the summer. It is not yet law, so the NIS Regulations 2018 remain the operative duty today. When it passes it is expected to widen the regime in ways that matter for anyone building AI on operational data:
- More organisations in scope, including managed service providers, data centre operators and designated critical suppliers, closing the supply-chain gap that AI dependencies sit squarely inside.
- Stronger incident-reporting duties and a two-tier penalty regime.
- A statutory footing for the Cyber Assessment Framework, so the benchmark you are already measured against gains legal force.
None of this reverses the current position. It confirms it. An operator that maps its AI estate to the CAF now is preparing for the statutory version of the same test, rather than waiting for a rule it can already see coming.
Where the real gap sits
In most operators the missing piece is a clear answer to a governance question rather than a security product: for each AI component that touches the essential service, who is the accountable owner, is it recorded as an in-scope asset, and where does it sit against the four CAF objectives. Where that ownership is absent, the model runs in a grey zone that belongs to data science by habit and to the regulated estate by law, and the two assumptions only surface during an incident or an assessment.
The constraint
The binding limit is accountable ownership of each AI component as an in-scope asset under the existing NIS duty. Buying model-security products without assigning that ownership and tracing the system through the CAF spends budget while leaving the exposure in place, because the deciding question is whether a named person is answerable for the model inside the framework, and a competent authority expects that answer regardless of what tools are deployed.
What a board should ask the operations and security leads
Three questions settle it. Which AI systems touch the essential service, and are they on the asset register that the NIS duty covers. Who owns each of them, and can that owner evidence the controls against the four CAF objectives. And what happens to the operation if the model is wrong or unavailable, which is the manual-fallback question the resilience objective already demands. An operator that can answer those is compliant on AI in the only way the regime recognises, and ready for the statutory framework when the Bill completes its passage.
Sources
- The Network and Information Systems Regulations 2018.
- National Cyber Security Centre, Cyber Assessment Framework, and CAF v4.0 released in response to growing threat, August 2025.
- UK Parliament, Cyber Security and Resilience (Network and Information Systems) Bill, introduced November 2025, in the House of Lords during 2026.