AI and the Cyber Security and Resilience Bill
The Cyber Security and Resilience Bill (CSRB) has no general AI regime, but the Lords added AI measures in October 2026: a review of AISI audit powers and new powers over AI agents. Many commercial AI services are likely to be in scope through existing categories.
The Bill is not yet law. Regulation and section references come from the Bill as brought from the Commons (HL Bill 32). The Lords has since amended the Bill, so numbering and detail may change in the final Act. This guide is general information, not legal advice.
Where this stands
The Cyber Security and Resilience Bill (CSRB) was not written as an AI law, and it has no general rules on building AI models. AI entered the debate in the Lords, and three things came out of Report Stage on 26 October 2026: a required review, new powers concerning AI agents, and a statement that many AI services are likely to be in scope.
Any of these could change. The Commons still has to consider Lords amendments, and the Government opposed one of them.
What happened at Report Stage
| Amendment | What it does | Result |
|---|---|---|
| 163 | Requires a review of whether the AI Security Institute (AISI) should have statutory pre-deployment audit powers | Agreed 202 to 155, a Government defeat |
| 70 | New powers where an incident involves loss of control of an AI agent | Agreed without a vote |
| 41 | A duty on developers to report incidents | Rejected 95 to 165 |
| 66 | AISI review of frontier AI products on essential-service networks | Not moved |
| 79 | AISI powers to set standards, inspect and require remedial action | Not moved |
These results come from our Report Stage summary, which relied on parliamentary records and monitoring reports.
What Amendment 163 would actually do
It does not give AISI any audit powers. It would require a review within 12 months of Royal Assent of whether AISI should be able to audit frontier AI systems before they are deployed on essential-service networks, followed by a report to Parliament.
The Minister, Baroness Lloyd of Effra, opposed it. She argued that a statutory role could undermine the voluntary access AISI has to models. The Government could seek to reverse it in the Commons.
What Amendment 70 means
Amendment 70 is a Government amendment, so it is more likely to survive. It gives the Secretary of State powers where a compromise involves loss of control of an AI agent. These reportedly include appointing inspectors, entering premises, running tests, requiring information held outside the UK, and imposing financial penalties.
If you run essential services using AI agents, this is the part to read closely in the final text.
Are AI companies in scope?
There is no standalone AI category. An AI business would be caught, if at all, through the usual routes:
- As a cloud computing service (RDSP). During the debate the Minister said commercial AI services delivered as software or platforms are "likely to be in scope" of the NIS regime. Check this against the definition, which requires on-demand, self-service access to scalable and elastic computing resources.
- As a managed service provider (RMSP). An AI operations business may be caught if it manages customer systems under contract, with access to systems the customer relies on.
- As a critical supplier. A regulator may designate a supplier to a regulated organisation. The Minister said a non-UK model developer could be designated.
- As a data centre operator. AI data centres are regulated on rated IT load, like any other data centre.
Micro and small enterprises are excluded from the RDSP and RMSP categories, but not from critical supplier designation. See AI companies and AI users for more.
If you use AI rather than build it
Using AI does not create a separate duty under the Bill. But if you are regulated, your existing duties still apply to the AI systems you rely on.
- Include AI tools in risk assessments. The duty to manage risks to systems you rely on has no AI exception.
- Inventory AI agents and what they can touch. Record permissions, data access and who owns each one.
- Decide how you would stop an agent. Test a shutdown, and know what breaks when you do.
- Update your incident plan. Include a scenario where an agent takes unexpected actions. Decide how this fits with the significant incident test.
- Review AI supplier contracts. Ask about incident notice, logs, audit rights and sub-processors.
- Track the final wording. Look particularly at how Amendment 70 defines an AI agent and loss of control.
What we do not know yet
- Whether Amendment 163 survives the Commons.
- The final wording and scope of Amendment 70.
- Whether regulators will publish guidance specific to AI.
- Whether a separate UK AI bill will overlap with these powers.
Check the timeline for the latest position.
Sources
Frequently asked questions
Not as a separate category. AI businesses and services can be caught by existing categories, including cloud services, managed services, data centres and critical suppliers.
No. It would require a review, within 12 months of Royal Assent, of whether AISI should get them. The Government opposed it, so its fate depends on the Commons.
A Government amendment, agreed without a vote, adding powers where a compromise involves loss of control of an AI agent. Check the final text for detail.
An amendment for a developer incident reporting duty was rejected. If you are regulated, AI-related incidents go through the normal incident test.