Britain's National Cyber Security Centre, the cyber arm of the GCHQ spy agency, has judged that AI-powered malware will highly likely be used in a ransomware attack on the country's critical national infrastructure before 2028. Ransomware is the crime where attackers lock a victim's files and demand a payment, usually in cryptocurrency, to unlock them. Its chief executive, Richard Horne, links the threat to AI's ability to uncover weaknesses in IT systems that have been "tolerated for far too long".
That cause is the part that applies to everyone. A weakness that is tolerated is usually a weakness that nobody has examined, and AI changes how much examining costs.
Looking was the expensive part
The clearest case is PHP, the language behind most websites. Two teams found its flaws separately, each working with AI models, and the Wiz researchers who published them rank them among the most severe ever found in PHP. Versions going back to 2012 had them. The code is public and many people have worked in it, so the flaws were not hidden. Finding them simply took more hours than anyone had spent.
That is how attackers have always chosen targets. A skilled person's time is the cost, so it goes where the payoff is biggest: banks, hospitals, large retailers. For example, a regional distributor with a stock system installed twelve years ago on a server in the back office has been safe because a week of specialist work against it would not pay. If a tool can check thousands of such systems overnight, the payoff no longer has to be large, because the cost is close to nothing.
The last time British hospitals were hit, the flaw was already known
In May 2017 the ransomware called WannaCry disrupted at least 80 of the 236 NHS trusts in England, and about 19,000 appointments were cancelled. Every infected organisation was running Windows that was unpatched or no longer supported, even though NHS Digital had told trusts in March and April to install the fix.
That attack needed a known weakness and a fix that nobody had installed. The attack the NCSC expects can use known flaws in old systems in the same way, but it no longer has to wait for one. The PHP flaws were found before any fix existed, and AI also shortens the time between a fix being published and an attacker using it, which is why firms have less time to install updates than they used to. An Australian spy agency gave the same warning about old technology in September.
Cheap break-ins make attackers careless
The NCSC expects a second consequence by 2028. When a way into a network is cheap to replace, attackers stop worrying about losing one. Most small organisations count on spotting an intruder, locking the account and being done. An intruder who can find another way in by lunch does not mind being spotted, and may not bother to hide.
The fridge attacks across Europe show how loud that can get. Supermarkets in Belgium, Luxembourg and Sweden reported fridges and freezers going wrong overnight, and one Swedish store said its freezers had been set to 50 degrees. Nothing reported so far ties them to AI, but they show what a break-in looks like when it reaches the machines on the shop floor.
Plan for the day someone gets in
Defenders have the same tools. In the PHP case the people who found the flaws reported them, so the fix came out as the details became public, although the most powerful models so far go to selected companies. A small firm can still ask the supplier of each old system who is checking it for flaws now.
The rest is about the moment after someone gets in. Keep one copy of your data that is not connected to your network. Write down how you would take orders and pay staff if the main system were down for three days. Keep the most valuable system apart from the rest. An old system is only as safe as the effort needed to look at it, and that effort is now close to nothing.