NCSC's 2028 Forecast: AI-Powered Attacks on UK Infrastructure Will Ride on Today's Legacy Estate
By David Nakamura | AI & Data News
A forecasting note issued recently by the United Kingdom's cyber agency turns the usual AI-security debate on its head. The National Cyber Security Centre (NCSC), the cyber arm of Government Communications Headquarters (GCHQ), does not expect 2028 to bring an exotic new class of malware so much as a cheaper, faster version of attacks that already work today. Its assessment: by around 2028, artificial-intelligence-driven malicious software and ransomware campaigns against UK critical infrastructure are a realistic prospect, and the organisations most likely to be hit are those still running unsupported, internet-visible legacy systems. For global AI and data observers, the message is less about frontier models than about the state of the technical debt they will be pointed at.
The assessment, broken into three threat vectors
The NCSC rates the likelihood of hostile actors adopting AI techniques as very high. In its framing, rising model capability compresses the cost, speed and skill requirements of intrusion across three areas:
- Large-scale ransomware operations — AI tooling that automates code assembly, packaging and target selection, letting fewer operators run campaigns that once required larger teams.
- Higher-precision social engineering — personalised, fluent, language-matched deception that raises the success rate of credential and help-desk manipulation.
- Accelerated vulnerability discovery and exploitation — machine-assisted code analysis that shortens the gap between a flaw existing and a flaw being weaponised in the wild.
Reporting on the review has underlined the target set: coverage by The Guardian describes AI-powered malware as "highly likely" to be used against power stations, airports and hospitals by 2028, while other summaries of the document point to operational technology such as grid switching, water infrastructure controls, production lines and heating and cooling plant. The master assessment is consistent with both: the risk is concentrated where old connectivity meets physical consequence.
Why the warning lands in this particular moment
In recent years, the capability curve of large models has outrun most institutional planning cycles. The NCSC notes that technology companies including Google, Anthropic and OpenAI keep shipping more capable systems, and that some of these models already demonstrate an ability to find software vulnerabilities and analyse code defects. Those features were built to help developers and defenders; the same techniques are now transferable to attackers.
The defender-side evidence is concrete and public. Google DeepMind has described how its vulnerability-hunting agent Big Sleep located real zero-day defects in widely used open-source software, and has since presented CodeMender, an agent aimed at finding and proposing fixes for code-security flaws. Where such workflows once demanded scarce expert time, they are becoming automatable — which is precisely the asymmetry the NCSC is asking organisations to price in.
What actually breaches organisations today
A second, quieter line in the warning is the one most likely to change board behaviour. The NCSC stresses that attackers rarely succeed because of technological novelty. Successful intrusions, it says, typically follow from known vulnerabilities left unpatched, misconfigurations, and old systems that never needed to be connected to the public internet. The common entry points remain traditional: exposed public network interfaces, phishing emails, abuse of legitimate accounts, and flawed remote access services.
That framing has a practical implication for anyone reading the 2028 forecast: AI is an amplifier of an existing failure mode, not a substitute for it. Reducing internet exposure and closing known gaps is therefore a hedge against tomorrow's tooling, and it is affordable now.
2025 as a dress rehearsal for AI-scaled campaigns
The United Kingdom already has a recent, expensive example of what happens when today's tactics are applied at volume. In April and May 2025, a wave of ransomware and social-engineering attacks swept UK retail: Marks & Spencer suspended online orders for roughly six weeks and told investors it expected the incident to cost around £300 million in lost operating profit, while the Co-operative Group reported £206 million in lost revenue and the compromise of member data on the scale of millions of records. Harrods confirmed it had also been targeted. None of those breaches required advanced technology; access brokers and deception were enough.
Set that baseline against the NCSC's 2028 scenario. If a manually run social-engineering campaign can cost one retailer a nine-figure sum, an automated and better-targeted version of the same playbook — with faster phishing, faster code generation and faster exploitation — raises both the frequency and the ceiling of losses. Timelines from the industry data reinforce the point: GreyNoise recorded a median of 11 days between a surge in exploitation activity and the public disclosure of the underlying flaw, and datasets summarised by Mondoo, citing Mandiant's M-Trends research, describe time-to-exploit collapsing from around 63 days in the 2018–2019 period to single-digit days, while remediation cycles in many organisations remain measured in weeks.
The numbers behind the readiness gap
| Indicator | Figure | What it signals | Source |
|---|---|---|---|
| AI-driven malware / ransomware against UK critical infrastructure | Realistic by around 2028 | Two-year preparation window, not a distant risk | NCSC assessment, reported Nov 2026 |
| National "Cyber Shield" first phase | £60 million | National, AI-assisted defensive capability build-out | UK government investment plan |
| National Exposure and Real-Time Vulnerability Centre (NERV) | £70 million | Continuous mapping of exposure across infrastructure and the wider economy | UK government investment plan |
| Combined value of the two new programmes | £130 million | A share of the roughly £600 million AI-risk package set out in October 2026 | Reuters; UK government announcements |
| Central government IT estimated as legacy | 28% | The scale of the estate the forecast is aimed at | UK Public Accounts Committee |
| Highest-risk "red rated" systems still unfunded | 21 of 72 | Intent exists; money and time lag behind it | UK Public Accounts Committee |
| Reported financial impact of the 2025 retail attacks | £300 million (M&S) and £206 million (Co-op) | Cost baseline that AI-scaled campaigns would multiply | Reuters; BBC News |
| Median lag between exploitation surge and disclosure | 11 days | Defenders increasingly learn after attackers have acted | GreyNoise, via Cybersecurity Dive |
The boardroom message from Richard Horne
Richard Horne, the NCSC's chief executive, links the coming AI era directly to historical underinvestment. New technology, he argues, is magnifying legacy weaknesses that many organisations have tolerated for years; unless businesses and public bodies strengthen the fundamentals and patch core weaknesses, they will be easily exploited by future attackers. Over the past few decades, he notes, the habit of building digital systems at the lowest possible cost has treated security as optional. In an AI era, that assumption no longer holds.
His two requests to industry leaders are recognisably governance ones rather than technical ones: invest with more urgency, and make security and system resilience a baseline requirement of operations instead of a lesson learned after an incident. The NCSC likewise recommends raising cybersecurity investment, cutting unnecessary internet connectivity, shrinking the external attack surface, and treating basic cyber security practices as the starting point of defence — with the explicit view that building a secure estate is a core strategic duty of senior management, not solely a technology function.
A pre-2028 checklist for critical infrastructure operators
Translated into work that can be scheduled this year, the guidance forms a sequence:
- Inventory and de-expose. Enumerate every internet-facing asset, then remove or close non-essential ones — starting with operational technology, remote-access services and forgotten edge devices.
- Fix the known. Drive a patch cadence tied to exploit likelihood rather than raw severity scores, and measure mean time to remediate against the days-scale exploitation window.
- Harden identity. Phishing-resistant authentication, privileged-account separation, and verified-callback procedures for help-desk and supplier requests, where social engineering does most of the work.
- Segment the plant. Apply the secure-connectivity logic of the NCSC's operational-technology guidance: no more production and control traffic than is strictly necessary, and no accidental internet exposure.
- Plan the legacy exit. Give every end-of-life system a named owner, a funded replacement route and a compensating-control package while it stays in service.
- Rehearse the outage. Test recovery from an encrypted estate and from a data-extortion scenario without network access, with executive decision-making included.
- Put it on the board. Set resilience targets as operational requirements with budget attached, and report exposure reduction as a tracked metric.
FAQ: what the NCSC warning does and does not say
Is the NCSC predicting that AI will invent new attacks by 2028?
No. The centre's position is that hostile actors are very likely to use AI to conduct existing attack patterns — ransomware, social engineering, vulnerability exploitation — at lower cost, greater speed and larger scale. The techniques that succeed today are the ones being accelerated.
Why are old systems singled out as the priority target?
Because they combine three properties attackers like: they are documented, they no longer receive vendor fixes, and they are frequently still reachable from the public internet. That is why the agency's advice begins with exposure reduction rather than with tools procurement.
Is the UK using AI on the defensive side too?
That is the stated intent. Cyber Shield is positioned by the NCSC as a national, collaborative approach to AI-assisted — including agentic — cyber defence, and NERV is designed to continuously discover exposure across critical infrastructure and the wider economy. Both programmes were announced with first-phase funding of £60 million and £70 million respectively, inside a broader UK AI-risk package reported at roughly £600 million.
What is Manchester's role?
A new artificial-intelligence cyber security centre of excellence is to be established in Manchester and operated by the NCSC, with the remit of building AI-security capability across government.
What should a non-UK reader take from a British warning?
The dependency chain is global. The same legacy connectivity, remote-access patterns and help-desk trust assumptions appear in utilities, logistics, health systems and manufacturers worldwide, and vendors shipping AI-assisted code-analysis tooling do not respect borders. The transferable point is sequencing: exposure reduction and patch discipline buy protection against AI-era attacks now, while model-specific defences are still being defined.
What to watch next
The near-term signals are procedural. Watch how quickly the National Exposure and Real-Time Vulnerability Centre moves from funding to live visibility of the estate, whether Cyber Shield's agentic defence architecture is documented in a way third-party operators can adopt, and how the Manchester centre of excellence staffs and distributes capability across government departments. Watch also the two metrics that decide whether this forecast becomes a near-miss or a breach count: the number of internet-exposed legacy and operational-technology assets removed, and the time organisations take from an exploited disclosure to a deployed fix. Around 2028 will not judge the UK by its models but by its housekeeping.
Where the master assessment is clear-eyed is in refusing to let artificial intelligence become an excuse for postponement. The NCSC's conclusion is that AI will shorten the distance between a flaw being found and a flaw being used, and that for operators still dependent on old technology and traditional architectures, upgrading systems, closing vulnerabilities and building resilient security is no longer a long-range plan but an immediate task. The infrastructure that decides the outcome is already in place; only its maintenance schedule stands between today and the next campaign.