UK cyber bill leaves AI safeguards unresolved for businesses

Britain has not committed to adding AI-specific safeguards to its Cyber Security and Resilience Bill, despite a ministerial promise to legislate where necessary, according to The Register – Security. Officials are considering provisions covering serious incidents, near misses and disruption involving AI. For businesses, these remain possible obligations, not settled requirements.

What happened

At this week’s Business, Innovation, Science and Trade Committee hearing, senior civil servants told MPs that amendments addressing AI risks were still under consideration. The distinction matters: identifying a bill as a possible legislative vehicle does not establish what it will require, which organisations it will cover or how compliance will be enforced.

Alexandra Jones, the director general responsible for AI and emerging-technologies policy at the Department for Business, Innovation, Science and Trade, described three areas of work. These concerned preparedness for AI, cyber and loss-of-control risks; the identification and reporting of serious incidents and near misses; and the containment of AI-driven disruption across organisations.

But Jones’s account stopped short of confirming amendments. Asked about legislative changes, she said policy was still being developed to determine whether the bill should be used. The uncertainty therefore concerns more than drafting: officials have not publicly settled whether this particular legislation is the instrument for the job.

Responsibility is also distributed across government. Jones told MPs that recent departmental changes had left the Department for Digital, Culture, Media and Sport responsible for the Cyber Security and Resilience (Network and Information Systems) Bill. AI minister Kanishka Narayan sits in the Cabinet Office. The official explaining AI policy, the minister promising action and the department responsible for the bill occupy different parts of Whitehall.

The background

Britain’s existing approach to AI regulation is decentralised and sector-specific. The Information Commissioner’s Office, for example, has published principles explaining how it will apply data-protection law to AI. The absence of a commitment to new AI provisions in this bill should not be confused with the absence of existing law governing AI use.

That distinction is particularly relevant to companies already assessing their data obligations. The UK Data Act rollout is a separate compliance question from whether Parliament adds AI-specific cyber-security provisions. Uncertainty about one proposed measure does not resolve obligations under another legal framework.

The international picture is uneven. The EU’s AI Act entered into force in August 2024, with powers taking effect in stages. The Register reports that America has no federal legislation specifically aimed at AI, although some states are introducing laws. Britain’s deliberations consequently sit between an EU-wide legislative framework and an American landscape that also varies by jurisdiction.

Can AI already spread like a computer worm?

The immediate backdrop includes research published by Anthropic on Monday. It examined whether AI models could copy themselves between devices and operate on compromised hardware. The company forecast that small models could be months away from functioning as self-replicating worms on vulnerable connected devices.

That is a forecast, not a demonstrated outbreak. As The Register explains, Anthropic extrapolated from improving model capabilities; it did not demonstrate a fully autonomous worm spreading across real internet-connected devices. This is an important evidential boundary. A projection can justify preparation without establishing that the projected capability already exists in the wild.

The scenario also differs from vulnerabilities in the software through which businesses use models. The risks of centralising AI access concern another part of the security problem. Treating all such issues as a single category of “AI risk” can obscure whether a proposed safeguard addresses vulnerable infrastructure, model behaviour or the consequences of an attack.

What people are saying

Narayan’s position is broader than a pledge to amend this bill. Announcing plans this week to strengthen national defences against AI risks, he said the government would use its available powers and legislate wherever necessary. He also emphasised scientific expertise, technological capabilities and institutions. That leaves room for measures outside legislation as well as for new law; it does not specify either a clause or a timetable.

Committee chair Liam Byrne challenged the gap between that political posture and officials’ language. When Jones described using the bill to consider changes, he argued that legislation must actually legislate. His objection was about commitment: a policy discussion cannot itself create the reporting or containment arrangements officials say they are examining.

Consultancy Booz Allen Hamilton offered a more immediate assessment of the technical threat. It warned that autonomous AI systems were already capable of finding and attacking critical operational technology and industrial equipment, potentially cutting access to water, power and other necessities. The source reports the consultancy’s warning, rather than a documented incident demonstrating that entire chain of events.

Earlier committee testimony addressed who should be trusted to manage the risks. Daniel Kokotajlo, executive director of the nonprofit AI Futures Project, advocated avoiding recursive self-improvement and imposing substantially greater transparency and regulation on developers. He presented that approach as the least objectionable available option, not a guarantee of safety.

Byrne summarised the researchers’ testimony as a severe indictment of industry safety culture, competitive incentives and the prospects for self-regulation. That was his characterisation of the evidence heard by the committee, not an independently established finding about every developer. Researchers also warned that international rules would encounter entrenched political differences, complicating proposals that depend on agreement across borders.

What happens next

The next consequential step would be the publication of actual amendments. The source provides no confirmed wording, introduction date or implementation timetable for AI-specific measures. Until those appear, businesses cannot reliably identify the proposed duties or distinguish government ambitions from provisions Parliament will be asked to approve.

Near-miss reporting is one especially important unresolved detail. Reporting arrangements limited to realised harm would differ materially from ones also covering events that nearly caused it. Jones explicitly included near misses in the work under consideration, but the account gives no definition, reporting threshold or recipient for such reports. Those details would determine much of the practical compliance burden.

Containment raises a different boundary question. Officials referred to disruption across organisations, rather than only within a single business. Any resulting provisions would need to establish whose responsibilities begin and end where several organisations are involved. The source does not say whether the government intends duties for developers, service providers, users or some combination of them.

For American companies with British operations or customers, the missing scope is as significant as the missing timetable. Nothing in the reported proposals establishes which overseas businesses might be covered. The practical distinction to preserve is between monitoring a possible UK obligation and treating a ministerial announcement as evidence that it already applies.