UK Warns of AI-Powered Ransomware Against Critical Infrastructure by 2028

Britain's top cyber agency has put a date on a fear many security teams already had. The forecast is about 2028, but the problem it describes is sitting in systems running today.

Today is November 5, 2026, and this matters right now because the UK's National Cyber Security Centre (NCSC) issued the warning alongside its annual review on 3 November. It assesses that a ransomware attack involving AI-powered malware against UK critical national infrastructure is highly likely by 2028. Critical national infrastructure means the essential services a country runs on, such as energy. Ransomware is malicious software that locks an organisation's systems and demands payment.

NCSC chief executive Dr Richard Horne aimed his message at leaders, not just IT teams. He challenged those who have "got away with tech on the cheap", warning they may have to explain service failures and business losses if they do not act.

Why AI changes the picture

AI does not need to invent new kinds of attack to be dangerous. The immediate worry is that it makes existing attacks faster, cheaper and easier to reach.

The NCSC says work that used to need specialist knowledge, such as writing code, studying system designs and finding security weaknesses, can increasingly be automated. Less skilled attackers can attempt more advanced operations, and experienced groups can do more at once. Weaknesses that once drew little attention may become easy to find with automated tools.

The agency still believes AI can improve security in the long run. But buying new tools without fixing the basics will leave large gaps.

This is not a sudden warning

Pressure on UK infrastructure already exists. In a June 2026 address, Horne said the NCSC had managed more than 200 cyber incidents affecting critical national infrastructure and its supporting ecosystem in the year to May. Around three-quarters were believed to have links to state actors.

Those figures cover incidents the agency handled in that period. They are not a count of every attempted attack.

What the research shows, and what it does not

The UK AI Security Institute tested seven models released between August 2024 and February 2026 in simulated corporate and industrial environments. At a fixed processing budget, average progress through a 32-step corporate attack rose from 1.7 steps for GPT-4o to 9.8 for Claude Opus 4.6. The best single run completed 22 steps.

That is a steep climb, but read it carefully:

• Results were much weaker in the industrial-control scenario, which needed more specialist work and memory across complex stages.
• Neither environment had active defenders blocking or slowing the agents.
• The work does not show that autonomous AI can reliably carry out a full attack on a working power station or other essential service.

So the direction is worrying, while the end point is not proven.

Why old systems make this hard

Infrastructure cannot be replaced like a laptop. The government's Energy Sector Cyber Security Strategy, published in May, admits parts of Britain's energy infrastructure were not designed for today's connected and decentralised world. It points to mixing legacy equipment with newer technology, and to a shortage of people skilled in both cybersecurity and engineering.

Every fix must also consider service availability and physical processes. A change that is routine on an office server can carry real operational consequences here.

The window to react is shrinking

The NCSC's earlier assessment through 2027 said AI would almost certainly shorten the time between a vulnerability being disclosed and being exploited. In some cases that gap was already down to days.

It also warned of a widening divide between organisations that keep pace and those that fall behind. That matters because your resilience also depends on your suppliers.

Separately, the NCSC warns of a "patch wave": AI-assisted discovery could expose years of technical debt, meaning old shortcuts and unfixed flaws, across commercial, proprietary and open-source software. That means more security updates, arriving more often.

What to do now

These steps come from the NCSC's published guidance, in the order most teams can act on them.

Step 1: Know your exposure and clear unsupported technology

Find what faces the internet and put it first for patching. Where a system can no longer get patches, plan to replace it or return it to vendor support.

Step 2: Prepare for faster patching before you need it

Set up a process for assessing urgent vulnerabilities, testing changes and approving deployment. Check that critical suppliers can move at a similar speed. Automatic updating may not suit some operational technology and safety-critical systems, so the goal is less avoidable delay without giving up safe operation.

Step 3: Assume someone gets in, and limit the damage

Restrict privileges, segment networks so a compromise cannot spread easily, and improve visibility across important systems. Give extra protection to administrative accounts and backup infrastructure, because losing either makes everything worse.

Step 4: Prove you can recover

Attackers often target backups, so protect both cloud and on-premises copies. Plan for restoring applications, identity systems and configurations, not just files. Remember that restoring encrypted systems does nothing about stolen data or later extortion. Then run exercises. The real question is not whether backups exist, but whether essential operations can actually come back.

Step 5: Put limits on your own AI agents

If you use AI defensively, decide how much autonomy it truly needs and what happens if it misbehaves. An assistant that recommends actions is a very different risk from one that can change production systems alone. Do not rely only on safeguards built into the model. Use environmental restrictions, oversight, activity logging and a way to stop the agent.

Conclusion

The 2028 date gives organisations a planning horizon, but the weaknesses behind it already exist. Boards should look for progress in plain operational terms: fewer exposed systems, tighter control of privileged access, stronger suppliers and tested recovery. AI may change the speed and economics of attacks. The outcome will depend on decisions made now about the technology already in use.

Merits

• Gives leaders a clear planning horizon instead of a vague threat.
• Focuses on practical controls most organisations already understand.
• Makes security a board-level operational issue, not only an IT one.
• Separates measured research progress from what remains unproven.

Demerits

• A forecast is a judgement, not a certainty, and it could prove early or late.
• Research results came from simulated environments without active defenders.
• Legacy replacement is slow, costly and operationally hard.
• More frequent patching adds pressure on already stretched teams.

Caution

This article is educational and summarizes published assessments in plain language. It contains no placeholder values, but any settings or process you adapt must fit your own environment. Verify claims against the original NCSC and government publications before relying on them, and test changes safely, especially on operational and safety-critical systems.

Frequently asked questions

• What did the NCSC warn about in November 2026? — It assessed that a ransomware attack involving AI-powered malware against UK critical national infrastructure is highly likely by 2028.
• What is critical national infrastructure? — It is the set of essential systems a country depends on, such as energy and other services the public relies on every day.
• Can AI already attack a power station by itself? — The research cited does not show that; tests were simulated, had no active defenders, and industrial-control results were much weaker.
• What is a patch wave? — It is the NCSC's term for a surge of security updates as AI-assisted discovery exposes old flaws across many kinds of software.
• Should critical systems patch automatically? — Not always, because the NCSC recognises automatic updating may not suit some operational technology and safety-critical systems.
• Are backups enough against ransomware? — No, because attackers target backups and restoring systems does not solve stolen data, so recovery must be protected and tested.
• How should organisations use AI agents for defence? — Limit their autonomy, log their activity, keep oversight and make sure they can be stopped.

Tags

#NCSC #Ransomware #CriticalInfrastructure #AISecurity #Cybersecurity #OperationalTechnology #PatchManagement #CyberResilience #UKCyber #LegacySystems