AI Cyberattacks on UK Critical Infrastructure ‘Highly Likely’ by 2028

LONDON, November 3, 2026 — Britain should prepare for AI cyberattacks on UK critical infrastructure within the next two years, the country’s National Cyber Security Centre (NCSC) has warned in its latest assessment. The agency judges such an attack to be highly likely by 2028, as artificial intelligence gives criminal and state-backed operators more speed and scale in finding weaknesses and exploiting systems.

The warning is not a prediction that a particular attack is imminent, nor does it mean artificial intelligence has made conventional cybercrime obsolete. Reporting by Reuters, The Guardian, the Financial Times and The Associated Press describes a more immediate concern: attackers can use AI to accelerate familiar methods, while organisations continue to leave known vulnerabilities, exposed services and weak access controls unresolved.

The NCSC, part of Britain’s signals-intelligence agency GCHQ, assesses that AI-enabled ransomware against critical national infrastructure is highly likely by 2028. The Financial Times reported that the agency had received 1,831 reports of cyberattacks affecting UK organisations in the year to August, including 80 incidents classified as nationally significant. Those figures show the range of activity confronting defenders, although they do not mean every report involved artificial intelligence or a successful breach.

Critical infrastructure includes the systems that keep electricity flowing, water treated and distributed, communications operating, food moving through supply chains and essential manufacturing running. Hospitals, transport networks and public services also depend on digital systems and suppliers that may be connected to larger networks. An intrusion into one organisation can therefore have effects beyond the first victim, particularly when a supplier, managed service provider or widely used software product sits between multiple operators.

NCSC chief executive Richard Horne has warned that underinvestment and outdated technology have left organisations with weaknesses that attackers can exploit at greater speed. In the assessment described by several news organisations on Tuesday, the agency’s central concern is not that AI has created an entirely new category of attack. It is that the technology can help hostile actors identify targets, search for exploitable flaws, tailor phishing messages and develop or adapt malicious code more efficiently.

The agency identifies China as Britain’s most sophisticated and sustained state cyber threat, while also warning about Russian activity against military networks and Iranian cyber operations. These are assessments of national capabilities and intent, not proof that any named government carried out a specific incident mentioned in this article. Attribution for individual attacks requires evidence about the tools, infrastructure, operators and intelligence surrounding each case.

The NCSC’s warning also covers criminal groups that use ransomware to extort organisations. Such groups do not need to invent a new technique to cause serious disruption. They can gain access through stolen credentials, unpatched software, insecure remote-access services or a supplier with weaker defences, then use automation to move more quickly through the victim’s systems.

Cybersecurity teams have long dealt with reconnaissance, social engineering, credential theft, vulnerability exploitation and attempts to move between connected systems. AI can make parts of that work faster. A tool can help an attacker sort information about a target, generate convincing messages for different employees, examine code for possible weaknesses or modify scripts for a particular environment. The usefulness of these tools depends on access, technical skill and the security controls protecting the target; AI does not automatically turn every attempt into a successful intrusion.

The distinction matters because it changes the time available to defenders. A vulnerability that remains unpatched for months may once have attracted little attention from opportunistic attackers. If automated systems can find and test weaknesses at scale, the same delay can become more dangerous. Security teams may have less time between the discovery of an exposed service and an attempt to exploit it, particularly where the organisation lacks reliable asset inventories, central monitoring or a tested process for applying emergency fixes.

AI can also assist defenders. Security researchers can use machine-learning tools to review code, prioritise alerts, identify unusual activity and help produce fixes. Britain’s government said on October 12 that frontier AI was being put to work strengthening cyber defences, building on pilot projects that had identified and addressed more than 1,000 vulnerabilities. The government also said it was considering stronger protections for emerging AI risks through the Cyber Security and Resilience Bill regime and was evaluating the latest generation of models through the AI Security Institute.

That defensive potential does not remove the need for basic controls. A model that identifies a flaw is useful only if an organisation can assess the finding, verify it and deploy a safe correction. Teams also need to prevent sensitive source code, credentials and customer data from being sent to tools that are not approved for that purpose. AI adoption without clear permissions and oversight can create new routes to exposure even as security teams use the same technology to improve detection.

The UK’s experience is relevant beyond Britain. Indian banks, telecom operators, government agencies and technology companies face many of the same challenges: extensive software supply chains, older systems that cannot be replaced quickly, internet-facing services and pressure to deploy AI before governance and security practices have caught up.

Also read: RBI Steps Up Scrutiny of Banks’ Cybersecurity Controls

A series of suspected cyberattacks on supermarket refrigeration systems across Europe offers a separate example of how an intrusion can affect physical operations. Reuters, The Guardian and The Associated Press reported on incidents in which refrigeration or freezer controls were manipulated in stores across several countries, causing temperature spikes and forcing some retailers to remove or discard chilled and frozen stock. No group had claimed responsibility in the reporting reviewed for this article, and investigators had not established whether all the incidents were connected.

Those refrigeration incidents should not be presented as evidence that AI was used. The available reports do not establish that link. Their relevance is that connected equipment can become a route to disruption when remote access, software maintenance and network separation are not handled securely. Refrigeration controllers are operational technology: they influence equipment in the physical world, and an unauthorised change can spoil food even when no customer database has been stolen.

The same principle applies to industrial control systems, building management networks, energy operations and parts of transport infrastructure. A security team may be able to restore a server from a backup, but restoring a physical process can require checks on machinery, product quality, safety conditions and the integrity of control settings. Operators may need to isolate equipment while they establish whether it is safe to reconnect.

That is why the boundary between information technology and operational technology matters. Office systems and industrial controls often have different uptime requirements and different replacement cycles. A patch that is routine for a laptop may require testing, a planned shutdown or approval from an equipment supplier when applied to a plant controller. Those constraints are real, but they cannot be allowed to become a permanent reason for leaving exposed systems unprotected. Operators need compensating controls, including network segmentation, restricted remote access and monitoring for unusual commands.

The supply-chain risk is equally important. A large operator may maintain strong internal controls but still depend on small contractors, software vendors or remote-maintenance providers. Attackers often look for the least protected route into a larger organisation. Security requirements therefore need to cover suppliers, account privileges, remote connections and the process for notifying customers when a vulnerability is found.

Also read: CERT-In Warns AI Agents Accessed Indian Organisations’ Data

The NCSC’s warning reinforces a set of defensive measures that are familiar but inconsistently implemented. Organisations should maintain an accurate inventory of internet-facing systems, identify which services are business-critical and prioritise security fixes according to exposure and potential harm. Software updates should be applied promptly, and systems that cannot be patched immediately should be isolated or protected by additional controls while a safe remediation plan is prepared.

Access management is another priority. Multi-factor authentication should protect administrator accounts and remote access, while privileges should be limited to what each person or service needs. Unused accounts, old test environments and obsolete remote-access tools should be removed. Organisations should rotate credentials that may have been exposed and avoid placing passwords, API keys or other secrets in public code repositories or unprotected configuration files.

Network separation can limit the damage if an attacker enters one part of an organisation. Office networks, production systems, backups and operational technology should not be able to reach one another without a clear business need. Remote access for suppliers should be time-limited, monitored and approved. For critical systems, teams should consider whether an internet connection is necessary at all and whether a gateway or controlled maintenance process can reduce exposure.

Backups need to be protected from the same attack that could compromise the main environment. Organisations should keep recoverable copies, restrict who can alter or delete them and test restoration rather than assuming a backup will work during a crisis. Incident exercises should include realistic decisions about service continuity, communications, legal obligations, supplier coordination and how to keep essential functions operating while systems are isolated.

Leaders also need to treat cyber resilience as an operational responsibility rather than a technical issue delegated entirely to an IT department. Boards and senior managers should understand which services would be affected by a prolonged outage, how quickly those services could be restored and which suppliers could prevent recovery. Budget decisions should reflect the consequences of downtime, not only the cost of security software or compliance work.

AI systems themselves require controls. Organisations should define which tools may access sensitive data, restrict agent permissions, log consequential actions and keep human approval for high-impact operations. Testing environments should be separated from production systems and should use synthetic data and limited credentials where possible. A model’s instruction to behave safely is not a substitute for technical restrictions that prevent unauthorised network access or actions.

The NCSC’s 2028 assessment is a warning about a credible and growing risk, not a statement that Britain’s power grid or water systems are certain to be attacked on a particular date. “Highly likely” is the agency’s assessment of the risk horizon reported by the news organisations, not a public attribution of an operation already under way. The public reporting reviewed here does not identify a specific target or disclose a plan for a future attack.

Nor does the warning establish that AI is the cause of every recent cyber incident. Many breaches continue to rely on known vulnerabilities, stolen passwords, poor configuration and social engineering. AI may make those methods more efficient, but the effectiveness of an attack still depends on the target’s exposure and the controls in place. Keeping that distinction clear helps organisations prioritise practical protections instead of treating AI as an unknowable threat.

For the public, the most immediate effect may be indirect: disruption to services, supply chains or businesses that people rely on. There is no basis in the reporting reviewed for this article to suggest that households should expect a specific outage. Individuals can nevertheless reduce common risks by enabling multi-factor authentication, installing updates, using unique passwords or passkeys where available, and being cautious about unexpected messages asking for credentials or urgent payments.

The next test will be whether the warning produces measurable changes in resilience. Useful indicators include faster patching of serious vulnerabilities, fewer exposed remote-management systems, improved recovery times, stronger supplier controls and better reporting of incidents that affect essential services. Government guidance and new legislation can set expectations, but they cannot substitute for organisations knowing what they operate and rehearsing how to recover when something goes wrong.

Further detail from the NCSC’s full assessment, government implementation plans and sector regulators will help clarify how the 2028 risk judgement is being translated into requirements for operators. Until then, the central message is straightforward: AI is increasing the pace of the contest, while many of the weaknesses it can exploit are already known. The organisations best placed to withstand an attack will be those that fix those weaknesses, limit the routes an intruder can take and practise restoring essential services before a crisis.