No genius required: why Europe’s buildings have become an easy target

Three weeks ago I wrote in public that the physical economy, from supermarkets to office towers, is wide open to attacks that need no genius, and that AI is about to make the genius cheap (1). Since then, a senior executive at one of the leading AI labs has put a timetable on it, the UK’s national cyber agency has confirmed the acceleration, and the supermarket fridge story I posted on LinkedIn has come back in a second wave across seven European countries and, reportedly, to the storage of blood plasma. This piece explains why this is suddenly serious, what has happened since, and why the hidden technology in buildings is the weak spot. The good news comes with it: most of what protects you is basic security hygiene known since the 1980s, and a new generation of companies is building the tools to apply it at scale, which makes this as much an opportunity as a threat. No technical background needed.

Eight things to take away

1. Old weak spots, new economics. The flaws in building technology have been there for years. AI makes finding and exploiting them fast and cheap, and attackers can do it from anywhere, with little risk of being caught. Why this is suddenly serious
2. Not AI’s fault, but AI’s speed. France’s cyber agency found that the fridge attackers needed no clever exploit: the controllers sat openly on the internet with weak passwords. Sloppy security was always the problem; AI just exposes it faster. Time to clean up, quickly. Why this is suddenly serious
3. Governments confirm the acceleration. An executive at a leading AI lab told companies to brace for about 18 months; two weeks later the UK’s national cyber agency warned that AI will speed up attacks on critical infrastructure. We use the 18 months as our planning horizon. Three weeks that turned a warning into evidence
4. The fridge attacks were only the start. A second wave hit supermarkets in seven European countries, and according to a French newspaper, equipment storing blood products was manipulated the same night. The fridge story
5. The hidden technology in buildings is the weak spot. A flaw found with AI sits inside products from Siemens, Schneider, Johnson Controls and others, and much old equipment has no update path at all. Why the hidden technology is the weak spot
6. 18 months for software, a decade for buildings, and money moves first. We expect lenders, insurers and buyers to start pricing exposed, unsupported technology in 2027, before any law forces owners to act. Why 18 months · Our forecast
7. The basics work, so start this week. Operators who took their controllers off the open internet after the first wave largely escaped the second. Eight simple questions for your technology and facilities leads will show where you stand. How to prepare · Eight questions
8. A market is opening. First the clean-up of Europe’s installed base, then a new generation of connected buildings that are secure by design: a large opportunity for start-ups and the investors behind them. Where the opportunity is

Why this is suddenly serious

In plain words: The weak spots in your buildings have been there for years. What is new is that finding and using them has become fast, cheap and low-risk.

Every modern building runs on software you never see. The controller that keeps a cold room at −20°C, the box that runs the heating, the panel that opens the doors, the gateway that reports what the solar roof produces: each is a small computer. Their software is often ten or twenty years old and built from the same shared components across whole industries.

All software contains mistakes. Some of them let a stranger take control from outside. Security people call such a mistake a vulnerability, and a piece of code that uses it to break in an exploit. Most vulnerabilities in building technology are not new. They were there when the equipment was installed, and they are still there today.

For a long time this rarely mattered, for a simple economic reason. Finding such a mistake took a rare specialist weeks of work, and turning it into a working attack took more. Few attackers spent that effort on the cold room of a supermarket. Obscurity was most of the protection.

AI has removed that protection. Today’s AI models can analyse large amounts of code far faster than human teams. In February, Anthropic reported that one of its models found more than 500 serious flaws in open-source software that had been reviewed and tested for years, some of them for decades (2). In a prepared lab test in October, the security firm Claroty had an ordinary AI coding assistant rebuild a working attack on known weaknesses of the fridge controller used in French supermarkets. With some human guidance, it took about 28 minutes and $6 of computing (3).

And the attacker does not have to be anywhere near the building. Many of these devices can be reached directly from the internet, and public search engines for connected devices list them by brand and model. An attack can be run at night from a laptop in another country, through rented servers. Three weeks after hundreds of French supermarkets lost their frozen stock, nobody has claimed the attack and nobody has been publicly named (4). For the attacker, the risk of being caught is low.

Think of it this way. Every lock in your portfolio comes from a handful of manufacturers, some of them decades ago. Until now, picking one of those locks needed a master locksmith standing at the door. AI has handed everyone a machine that designs a key for many of these locks in half an hour, and the internet lets them try it on every door in Europe without leaving the sofa.

Fifteen years ago, an attack of this kind was the work of states. Stuxnet, which the New York Times reported to be a joint US-Israeli operation, took years of preparation to manipulate the Siemens controllers that ran Iran’s uranium centrifuges at Natanz. It drove the machines to destruction while playing back recordings of normal operation, so that the operators saw nothing wrong (5). It was the most sophisticated cyber weapon the world had seen. The supermarket attacks follow the same pattern: change what a controller does while the alarms stay quiet. What once took the world’s best intelligence services is now within reach of someone on a sofa, between lunch and dinner.

Before anyone blames AI, though, one thing needs saying clearly. Almost nothing in this story is new. France’s cyber agency ANSSI told the Senate that the fridge attackers did not even use the weaknesses researchers had published. They exploited elementary failures: controllers exposed directly on the internet, administered without protection, with weak passwords. ANSSI has advised against exposing industrial systems directly to the internet since 2014 (6). The wider rules are older still: do not leave systems reachable that do not need to be, change the factory passwords, keep software up to date, separate the systems that run a building from everything else, know what you own. They have been taught since the 1980s. Sloppy security was always the problem; it just stayed hidden because finding it was expensive. AI is not the burglar here. It is the auditor who arrived years early, and is now showing everyone, attackers included, where the doors were left open.

That is also the good news: the first steps are well known, cheap and need no new invention. It is time to clean up, and to do it quickly.

Three weeks that turned a warning into evidence

In plain words: An AI lab warned, a flaw found with AI put a large part of the internet at risk within days, and a government agency then confirmed the whole picture officially.

The warning. On 22 October, Jason Clinton, deputy chief information security officer at Anthropic (the CISO is the executive responsible for protecting a company’s systems and data), published a short post on X that has since been viewed around three million times (7). The past days had been bad, he wrote, and the coming months would be worse. Security leaders should ask their boards for help, “not for AI, for the basics”, and hold on for roughly 18 months. Salesforce CEO Marc Benioff replied: code red (8). His advice was not to buy AI, but to fund the unglamorous work of fixing, inventorying and isolating systems, as he had already urged in August (9).

The test case. The same week, two critical flaws surfaced in PHP, the programming language behind roughly 70 percent of all websites (10) and, as we will see, behind the control panels of many devices. They had been found with the help of AI by the security firm Calif and, independently, by Wiz (11). Details leaked early, and Amazon saw first exploitation attempts four days before the fix, called a patch, was released on 21 October (12). Germany’s federal cyber security agency BSI raised its warning to level 3 of 4 and advised operators to take servers offline if they could not install the fix quickly (13). Mainz switched off its city portal; the websites of Berlin’s police and Braunschweig’s municipal hospital went offline (14). Japan’s prime minister and the US NSA and FBI issued similar calls (15, 16). In a separate incident the same week, one issuer of the digital certificates that prove a website is genuine was found to use secret keys so weak that one of them was the number 4 (17, 18).

The confirmation. On 3 November, the UK’s National Cyber Security Centre (NCSC) published its annual review (19). It expects AI to act as a force multiplier: not new kinds of attack, but existing ones at far greater scale and speed, forcing everyone into a race to fix faster. It judges it highly likely that AI-assisted malicious software will be used in a ransomware attack (one that locks systems until a ransom is paid) on UK critical infrastructure by 2028. And it is spending money on the answer: £70m for a national centre that tracks exposed weaknesses and £60m for Cyber Shield, a national network of defensive AI agents (20). A test of AI against the government’s own networks found hundreds of potential problems (21).

The spread. By then the PHP flaws were being used in real attacks (22), and manufacturers of industrial and building equipment, including Schneider Electric, Siemens, ABB and Johnson Controls, had confirmed that specific products and versions were affected, some still without a fix (23). Attackers, meanwhile, are automating: Microsoft linked one attack to a group that let AI agents run everything from scouting to data theft, in some cases gaining privileged access within three hours (24), and the industrial security firm Frenos found that three of the six frontier AI models it tested could draw up workable attack plans against industrial control systems in simulated scenarios (25).

The fridge story, three weeks on: from groceries to blood

In plain words: What looked like a prank on supermarket freezers has come back in a second wave across seven European countries and, reportedly, reached equipment storing blood products. The same equipment runs cold rooms in logistics, pharmacies and hospitals.

When I posted about the French supermarkets in October, many readers took it as a curiosity: hackers warming up yoghurt. Three weeks later it reads differently.

First, how these systems work. A supermarket’s refrigeration has a small control computer for each display case and cold room, and above them one supervisory controller per store that sets the targets: temperatures, defrost cycles, alarms, and the temperature log food inspectors ask for. Its web page is how a refrigeration contractor fixes a fault at 2 a.m. without driving to the store. That web page is what the attackers used. They needed no malicious software and demanded nothing. They simply told the controller to do what it was built to do: defrost, and keep defrosting, while alarms waited.

Three lessons from the second wave matter far beyond retail.

It is a pattern, not an incident. The same families of controllers, the same overnight timing, the same manipulation of temperatures and defrost cycles. Still no claim, no ransom and no attribution. Researchers think the target lists were built simply by searching the internet for the controllers’ identifying features, such as product names and login pages (4). Anyone with a laptop can repeat that.

The basics work. The Belgian contractor Sabcobel, which runs refrigeration for Delhaize, said it knew of about 20 affected units out of the hundreds it manages. Most of its customers had taken their controllers off the open internet and onto a secure connection after the French attack; by its account, the units hit were largely those not yet moved (31, 33). Several affected units now need physical rework, not just a software update. Taking a device off the open internet costs little. Leaving it there put a piece of the food supply at risk.

Medical cold chains are in the line of fire. Blood products, vaccines and laboratory samples sit in the same kind of cold rooms, often run by the same contractors. A spoiled supermarket freezer is more than a write-off: food is a critical supply chain, and empty shelves unsettle people fast, as the pandemic showed. A spoiled plasma store would cost treatments. In the reported EFS case, no loss has been made public, and a technician watching temperatures caught the rise: the kind of independent check the controller itself cannot provide (29).

The law has not moved. Supermarkets generally fall outside NIS2, the EU’s main cybersecurity law for essential services (34). Hospitals and blood services fall under it, but the contractors who run their refrigeration often do not. And the EU Cyber Resilience Act, the new law for connected products, obliges manufacturers to report attacks on old controllers, not to fix them. That is by design, not a loophole: the law sets rules for products going forward and does not reach back to equipment already sold and out of support. Fixing the installed base stays the owner’s job (35).

Why the hidden technology in your buildings is the weak spot

In plain words: Your IT department routinely patches laptops and servers. The boxes in the plant room, however, are rarely patched, and some never can be.

Most coverage of PHP2Shell was about websites. For owners of real assets, the more important story sits elsewhere. PHP does not only run web shops and city portals. It also runs the small web pages through which technicians configure firewalls (the devices that guard the entrance to a network), storage devices, building management systems and access control panels. Johnson Controls, one of the world’s largest suppliers of building management and access control, has products on the list of affected vendors (23).

Then comes the uncomfortable number. In the customer data of the security firm Qualys, 71 percent of more than 150 million PHP installations with a known version run versions outside regular support. For devices built on such versions, any fix depends entirely on the manufacturer (36). Updating the company’s website does nothing for the firewall’s admin page, the storage box in the plant room or the controller behind the car-park barrier.

Four features make building technology one of the most exposed and most overlooked targets in the economy:

1. It is built once and rarely touched again. The software built into a device, called firmware, is assembled from shared components that are already old when the device ships. A study of popular routers used in industrial settings by Forescout and Finite State found components on average five and a half years old and 161 known weaknesses per device, 24 of them critical (37). The devices then outlive their support: the older Danfoss controller generation hit in France was declared end-of-life, meaning no more updates, in 2021 (30). This is code that has rarely been audited, and that AI now audits for a few dollars.
2. It is not treated as critical. Refrigeration, heating, ventilation and lifts are bought by facility management, serviced by mechanical contractors and judged on uptime and energy use. Maintenance contracts cover compressors and refrigerant, not software. The EU certification for refrigeration technicians covers refrigerants, leak checks and environmental law, but not network security (38). The result is a device that gets the attention of a boiler while carrying the risk of a web server.
3. Its connections are often invisible. For years, getting a controller online meant a local installer setting up port forwarding on the store’s internet router, a setting that makes a device directly reachable from the internet, exactly as the manufacturers’ manuals described (30). Often, nobody watched those connections afterwards. The shortcut has since moved to mobile networks: equipment ships with its own SIM card or a small 4G router that connects straight to the manufacturer’s cloud, bypassing the owner’s network entirely. In our own Berlin office, a WMF coffee machine suddenly started receiving alerts from the mobile network. That is how we found out it had a working mobile modem built in, although we never booked a connectivity option. It is no security flaw at WMF and harmless for coffee. But it shows that many devices today are always on without anyone having consciously ordered it, and a cold room is not a coffee machine. A connection that bypasses your network will not show up in your IT department’s monitoring unless someone lists it separately.
4. It belongs to nobody in particular. The manufacturer built the remote access, the installer set it up and holds the passwords, often the same ones for many sites, the operator owns the equipment, the landlord owns the building, and the IT department never saw any of it. When something goes wrong, every party can point to another.

That is why we expect the fridge nights to be the first of many.