Agentic AI resilience measures: what the Government plans under the Cyber Security and Resilience Bill

On 3 November 2026 the Government published a factsheet setting out four measures it plans to introduce for agentic AI "loss of control" incidents under the Cyber Security and Resilience Bill (CSRB). None of them is law. Each needs the Bill to receive Royal Assent, then a public consultation, then regulations approved by both Houses of Parliament.

Position (9 November 2026): The Bill has passed both Houses and is back in the House of Commons for consideration of Lords amendments. No date has been set for that stage. The power the measures rely on, clause 32 of the Bill as amended on Report (HL Bill 62), was added by the Government at Lords Report Stage on 26 October 2026. No consultation on the agentic AI measures has been published, and the Bill is not yet law.

What did the Government publish on 3 November 2026?

The Government updated its Bill factsheets on 3 November 2026; GOV.UK now lists them under the Department for Digital, Culture, Media and Sport (DCMS) as well as DSIT, which closed on 2 November 2026. We read this as DCMS taking over the cyber security brief, but the Government has not said so in terms. The change note reads: "Updated existing factsheets and added 2 new ones, 'Agentic AI resilience measures' and 'Vendor risk management'."

The Agentic AI resilience measures factsheet says the Government "will consult on using these regulation-making powers to introduce measures to help regulated entities manage agentic loss of control incidents in relation to their networks and information systems."

It sets a deliberately narrow scope. The Government's intention "is for these measures to cover incidents where operators are unable to adequately monitor or control AI agents", and the measures "are not intended to establish a comprehensive framework for AI governance or to tackle general AI harms."

What are the four planned agentic AI measures?

The factsheet lists four measures the Government "intends to introduce ... in the future, following consultation". The right-hand column is our reading of which part of HL Bill 62 each would rely on; the factsheet itself refers only to "the delegated power in the Bill".

#	Planned measure	What the factsheet says it would do	Power in HL Bill 62 (our reading)
1	Security and resilience requirements	Technical and organisational measures that reduce the likelihood and impact of in-scope incidents caused by AI agents, for example "technical measures for monitoring agents' actions or agent risk assessments"	cl 29(1) regulations imposing requirements on regulated persons (cl 30)
2	Affected party notification	Entities that have reported an incident caused by AI agents may have to inform other organisations those agents have affected, going beyond the Bill's customer notification duty	cl 29(1) and cl 30 requirements, or cl 32(5)
3	Evidence preservation	Entities that have reported an incident would preserve information on how AI agents could have caused it, "for example actions taken by agents, or operator attempts to control them"	cl 32(5)(b): "obtaining, preserving and analysing information"
4	Post-incident review powers for Ministers	Ministers could task technical experts in government to review agentic loss of control incidents, request information and run tests, then share lessons with regulated entities	cl 32(1), (5) and (6): functions for the Secretary of State, including inspectors and tests

The factsheet's worked example describes an operator whose AI agents start accessing and altering another organisation's database. The operator reports the incident, tells the database owner, and keeps records of the agents' actions and reasoning (it gives "chain of thought" logs as an example). Government experts then review the evidence, advise the operator and later share an incident review with the sector.

Which clause of the Bill do the measures rely on?

They rely on clause 32 of HL Bill 62, headed "Functions of the Secretary of State: loss of control of artificial intelligence agent". The Government inserted it at Lords Report Stage on 26 October 2026 as amendment 70, which the Lords agreed without a division. Related Government amendments 74 and 75 (penalties) and 87 to 89 (consultation and procedure) were agreed alongside it.

Clause 32 would let regulations under clause 29(1) confer functions on the Secretary of State where a security or operational compromise "is or may be attributable to loss of oversight or control of an AI agent", or where there is a risk of one. Those functions can include:

• identifying and investigating AI-related compromises
• obtaining, preserving and analysing information
• assessing causes, effects and risks
• testing and verifying remedies and mitigations, and sharing what works
• appointing inspectors who can enter and search premises, seize documents, require information (including information held outside the UK), carry out tests and interviews, and require steps to be taken (clause 32(6))
• sanctions, including financial penalties, for non-compliance (clause 32(9))

Two points limit it in practice:

1. "AI agent" and "loss of oversight or control" are not defined in the Bill. Both are left to regulations (clause 32(2)), which may define them by functionalities, characteristics or capabilities.
2. It needs consultation and Parliamentary approval. Clause 43 requires consultation before regulations defining those terms (clause 43(1)(d)) or conferring clause 32 functions (clause 43(2)(d)), and those regulations need a draft approved by both Houses (clause 43(5)).

Penalties under Part 3 regulations are capped by clause 33: for an undertaking, the greater of £17 million and 10% of worldwide turnover, and £17 million otherwise. That is the ceiling Parliament would allow, not a figure the Government has proposed for agentic AI.

What did the minister say in the Lords?

Moving the AI amendments at Report Stage, the Lords minister, Baroness Lloyd of Effra, told the House that "Amendment 70 expands this to clarify that future regulations may enable the Government to use in-house expertise, such as AISI, and gather information on why agentic incidents in our essential services happen."

She went on: "I will also bring forward three further measures in secondary legislation to manage the risks posed by agentic AI. These will be consulted on." She listed security and resilience requirements, preservation of information on incidents that stem from agents malfunctioning, and a requirement to tell other organisations if agents are negatively affecting their operations.

The factsheet sets out the same package: the post-incident review function linked to amendment 70, plus the three measures she listed in the Lords.

What is consultation and what would be law?

Item	Status on 9 November 2026	Source
Clause 32 power	In the Bill as passed by the Lords (HL Bill 62). Awaiting Commons consideration of Lords amendments and Royal Assent	HL Bill 62; Bills API
When clause 32 would commence	On Royal Assent: it sits in Chapter 3 of Part 3, which clause 69(1)(b) brings into force on the day the Act is passed	HL Bill 62 cl 69
The four agentic AI measures	Planned. Government "intends to introduce" them "following consultation"	GOV.UK factsheet, 3 Nov 2026
Consultation	Not published. We found no DSIT or DCMS consultation on GOV.UK as at 9 November 2026	GOV.UK search, 9 Nov 2026
Definitions of "AI agent" and "loss of control"	Not drafted. To be developed "with industry, oversight authorities, and other relevant organisations"	Factsheet; HL Bill 62 cl 32(2)
Regulations	None laid. Each measure would need the affirmative procedure	Factsheet; HL Bill 62 cl 43(5)
Start date	No date has been set. The factsheet says all measures "will be brought into force through secondary legislation following Royal Assent"	Factsheet

The power would exist from Royal Assent, but it has no practical effect until regulations are made under it.

How does this relate to the rest of the AI debate in the Lords?

At the same sitting the Lords rejected amendment 41, which would have required frontier AI developers to report incidents, by 95 votes to 165 (division 3729). They then agreed amendment 163, now clause 66, by 202 to 155 against the Government (division 3730). Clause 66 requires a review of the NIS regime within 12 months of Royal Assent, covering whether "frontier artificial intelligence models and systems, or other frontier agentic technologies" should be brought into scope and what powers the AI Security Institute should have.

The agentic AI measures are a separate route. They would apply to organisations already regulated, rather than to AI developers as such. Our guide to AI and the Bill sets out all three AI provisions side by side.

What this means: our analysis

This section is Precursor Security's interpretation, not a statement of the law.

• The target is the deployer, not the model maker. Every measure attaches to "regulated entities". If you run AI agents inside an essential, digital or managed service, the measures are aimed at you. If you only build models, they reach you only if you are regulated in your own right, for example as a cloud computing service.
• Logging is the measure most likely to bite first in practice. Evidence preservation only works if the evidence exists. Organisations that cannot reconstruct what an agent did, with what permissions and why, would struggle with measures 2 to 4.
• The definitions will decide scope. Until regulations define "AI agent" and "loss of oversight or control", nobody can say for certain which automations are covered. A narrow definition could leave scripted automation untouched; a broad one could reach it.
• Expect to be asked first. The Bill requires consultation and the affirmative procedure, so there will be a published proposal before any duty applies. That consultation is the moment to influence thresholds and logging requirements.

What should in-scope organisations do now?

1. Inventory your AI agents. Record each agent's owner, purpose, permissions, the systems and data it can reach, and whether it can act without human approval.
2. Make agent actions reconstructable. Keep logs of actions taken, tool calls, permission changes and operator interventions, with retention that would survive an incident investigation.
3. Test your stop controls. Know how to pause, isolate or revoke an agent, and rehearse it. The minister said existing direction powers could require "isolating systems, restricting data flows or suspending automated actions" in a loss of control scenario.
4. Add an agent scenario to incident response. Include who you would tell outside your customer base if an agent affected a third party's systems. See our incident reporting guide for the Bill's 24-hour and 72-hour notifications.
5. Watch for the consultation. Respond if you deploy agents in an essential or digital service. Our secondary legislation tracker records when it appears.
6. Check you are in scope at all. Start with the scope checker and the AI companies and AI users page.

Frequently asked questions

Are the agentic AI resilience measures law?

No. They are plans set out in a GOV.UK factsheet. They would be made as regulations under clause 32 and clause 29(1) of the Bill, after consultation and approval by both Houses, and only after the Bill receives Royal Assent. The Bill itself is not yet law.

What are the four agentic AI measures?

Security and resilience requirements for agentic loss of control; a duty to notify other organisations affected by an AI agent incident; a duty to preserve evidence about how AI agents caused a reported incident; and powers for Ministers to have government technical experts review such incidents, request information and run tests.

Which clause of the Bill covers AI agents?

Clause 32 of HL Bill 62, "Functions of the Secretary of State: loss of control of artificial intelligence agent". The Government added it as amendment 70 at Lords Report Stage on 26 October 2026, without a division. Clause numbers may change in the final Act.

When will the agentic AI measures apply?

No date has been set. The Government says all the measures will be brought into force through secondary legislation after Royal Assent, following consultation. No consultation had been published as at 9 November 2026.

Do the measures apply to AI developers?

Not directly. They are aimed at regulated entities that use AI agents in their networks and information systems. A developer would be covered only if it is regulated in its own right, for example as a cloud computing service or a designated critical supplier. A separate Lords amendment that would have placed incident reporting duties on frontier AI developers was defeated 95 to 165.

Sources

• GOV.UK, Agentic AI resilience measures factsheet, published 3 November 2026
• GOV.UK, Cyber Security and Resilience (Network and Information Systems) Bill: factsheets (change history, 3 November 2026)
• GOV.UK, Cyber Security and Resilience Bill collection
• UK Parliament, HL Bill 62, as amended on Report, clauses 29, 30, 32, 33, 43, 66 and 69
• UK Parliament Bills API, Lords Report Stage amendments (amendments 70, 74, 75 and 87 to 89 agreed)
• Lords Hansard, Report Stage, 26 October 2026 (Baroness Lloyd of Effra on amendment 70 and the three further measures)
• UK Parliament, Lords divisions 3729 and 3730, 26 October 2026
• UK Parliament, Bill stages

Position as at 9 November 2026. Clause numbers are from HL Bill 62 and may change in the Act.