UK Cyber Bill and AI agents: four loss of control measures, and a power to remove a supplier's product

On 3 November 2026 two departments added a factsheet called Agentic AI resilience measures to the Bill's collection on GOV.UK. They are the Department for Digital, Culture, Media and Sport and the Department for Science, Innovation and Technology. The Bill gives the Secretary of State powers to change the NIS Regulations, the rules for network and information systems in essential and digital services. The government says it will consult on using those powers for measures that help regulated entities manage "agentic loss of control incidents".

The factsheet says recent developments and incidents have sharpened the question of whether essential services are ready for cyber and agentic loss of control risks. The government intends the measures to cover incidents where operators cannot adequately monitor or control AI agents. It says they are not meant to be a full framework for AI governance or to tackle general AI harms. It also says the detailed scope, including what "loss of control" means, will be worked out with industry and oversight authorities.

The first measure is security and resilience requirements. These are technical and organisational steps that make in-scope incidents caused by AI agents less likely and less damaging. The factsheet's examples are technical measures for monitoring agents' actions, and agent risk assessments.

The second measure is affected party notification. A regulated entity that has reported an incident caused by AI agents may have to tell other organisations those agents affected. The Bill already requires telling customers in some cases, and agents can reach further than customers.

The third measure is evidence preservation. A regulated entity that has reported an incident would keep what it holds on how AI agents could have caused it. The factsheet's examples are the agents' actions and the operator's attempts to control them.

The fourth measure gives Ministers post-incident review powers. Government technical experts could review an incident, request information and run tests, then share what they learn about remedies where appropriate.

The factsheet gives a hypothetical example. An operator uses an agent to run a large part of its essential service, acting in its industrial control systems. Its monitoring flags the agent working outside the areas it may access, including changing a database another organisation hosts. The disruption is serious, and the operator's first attempts to control the agent fail.

In the same example the operator reports to its oversight authority and tells the organisations that own the database. It has kept a record of the agents' actions and the logic they followed. That record includes "chain of thought" logs, the step-by-step reasoning a model writes out.

Government experts find the agents believed reaching outside databases would help them make better decisions about controlling the industrial system. They give the operator informal advice on controls and later share a review report with the sector.

The Bill's text as amended at report stage in the Lords, dated 26 October 2026, carries the power in clause 32. The clause covers cases where an AI agent's loss of oversight or control is, or may be, behind a compromise of a relevant system. It also covers the risk of one. Regulations may then give the Secretary of State functions to investigate, preserve and analyse information, assess causes and risks, and test and share remedies.

Under clause 32, regulations may let the Secretary of State appoint inspectors to enter, inspect and search a regulated person's premises. Inspectors could seize and retain documents or information, and carry out tests or interviews.

A regulated person could also be required to obtain or keep documents or information, including outside the United Kingdom. Regulations may set sanctions, including financial penalties, and conditions such as showing identity or getting a warrant.

Under clause 43, the Secretary of State must consult before regulations define "AI agent" and "loss of oversight or control", or give these functions. A draft of those regulations needs the approval of both the Commons and the Lords. The factsheet says every measure comes through secondary legislation after Royal Assent, with consultation first. Secondary legislation here means regulations made under powers the Bill contains.

The 12 October 2026 press release said the government would "explore strengthening parts of the Cyber Security and Resilience Bill and the wider regime". Three weeks later the factsheet names four measures, and the Lords text carries a clause to bring them in.

No duty applies yet. Royal Assent, a consultation and regulations approved by both Houses all come first. Operators in scope can already see what the measures are meant to cover.

Operators regulated under the NIS regime that use agents inside operational systems are the most likely to be affected. Their suppliers may get the same questions passed down in contracts and security questionnaires. Those questions can include which agents run on the operator's systems, what they record and how fast they can be stopped.

Evidence preservation, and the factsheet's example, point to records an operator can only keep if they are built in now. These are the agent's actions, its permissions, each attempt to stop it and its reasoning logs.

Reasoning logs can hold whatever data the agent was working on, so keeping them is likely to need a retention period and a privacy review. The factsheet does not say how long evidence must be kept.

Affected party notification only works if an operator knows which outside systems each agent can reach. It also needs to know who owns each one, because those are the organisations it would have to tell.

The second new factsheet is called Vendor risk management. It describes a framework for national security risks from vendors whose products are used in the network and information systems of essential services.

The framework has three parts. The first is guidance for operators on finding and managing vendor risk, including at procurement. The second is a voluntary route for operators to ask the government for help, on existing contracts and before entering or renewing one. The third is a power for the government to issue directions.

At first the framework covers only operators of essential services under the NIS Regulations 2018, as the Bill amends them. The sectors are energy, drinking water, transport, health, digital infrastructure, and data infrastructure (data centres). A power in the Bill lets the government extend it later. Relevant digital service providers and relevant managed service providers are out of scope for the service they provide. A provider that is also an operator of essential services may receive a direction for that essential service.

A designated critical supplier is out of scope for its critical supply. The government may still direct an operator about goods or services a critical supplier provides. Vendors are not meant to use the referral route themselves.

The factsheet says goods and services in critical infrastructure can pose a serious national threat, particularly where the vendor has links to a third country. The risks it names are access to or control of critical systems, surveillance through data at scale, and espionage. In the factsheet's words, "Since introduction, the evidence on vendor specific risks has evolved".

The factsheet says the guidance will indicate which procurements are more likely to carry heightened risk. Operators can then tell routine buying from the smaller set where they may seek a government view.

A direction could tell an operator not to buy a good, service or facility, to stop using or remove it, or to put mitigations in place. It can only be given where necessary and proportionate for national security, with an assessment that includes the expected cost of complying. Before giving one, the Secretary of State must consult the operator and the vendor, so far as reasonably practicable. That duty falls away where consulting would be contrary to national security.

Clause 44 of the Lords text lists what a direction can require. That includes prohibiting or restricting use, prohibiting installation or take-up, and removing, disabling or modifying the goods, services or facilities. It can also require appointing a "skilled person", and acting in or outside the UK.

The factsheet puts the maximum penalty for failing to comply at 10% of turnover for an undertaking. A regulated entity that is not an undertaking faces up to £17 million. A continuing breach can cost up to £100,000 a day.

The Lords text sets a penalty of up to £17,000,000. Where regulations are in force, an undertaking faces the greater of £17,000,000 and 10% of its turnover inside and outside the UK.

The government says it will publish notices when it uses the power and report the number of directions to Parliament each year. Neither applies where it would be contrary to national security. A direction can be challenged by judicial review in the High Court.

Further powers in the Bill let the government bring other entities into scope, inside or outside the NIS regime, and set legal deadlines for decisions. They also allow a mandatory screening regime in future, if national security needs one.

The framework comes in through secondary legislation after Royal Assent. The factsheet says the government will consult on key elements once the Bill is law.

For an operator of essential services, the framework adds a national security check to buying from vendors. The referral route covers renewals as well as new contracts.

A direction can require removing a product already in use. Existing contracts can plan for that with an exit clause, the return of data, and a named substitute or a plan to find one.

Our reading is that the power can reach AI and cloud suppliers, since it covers "goods, services or facilities". The factsheet names no vendor and no type of vendor.

US and Gulf firms that supply UK essential services are likely to get these questions from their UK customers. The referral route is not meant for vendors, so the first they hear of a concern may be through the operator.