Hackers Turned Supermarket Fridges Up to 50°C in Sweden, Belgium, Luxembourg and Estonia.
Fridges in four countries were turned up on 2 November, one Swedish store to 50°C. Nothing was breached. The control panels answer to anyone.
On November 2 shoppers across Sweden, Belgium, Luxembourg and Estonia walked past supermarket refrigerators running warm. At an ICA Kvantum in Sjöbo, southern Sweden, the temperature setpoint had been moved to 50°C. Not switched off. Turned up, to roughly the temperature of a hot day in Death Valley, inside the cabinet holding the dairy.
Delhaize called it an "IT incident" and it reached dozens of its stores. Carrefour counted 15 affected out of more than 700. Coop reported six. The pattern repeated across four countries in a single day, six weeks after the same thing happened across France.
50°C
Where one Swedish store's refrigeration setpoint was moved to.
4
Countries hit on the same day. Sweden, Belgium, Luxembourg, Estonia.
2nd
Wave in six weeks. France went first, in October.
Nothing was breached
That is the part worth slowing down on. There is no zero-day here, no stolen credential, no clever chain of exploits. Refrigeration in a modern supermarket is run by a small building-management controller with a web interface, and a great many of those interfaces answer to anyone who types their address. An operator who works on these systems described the two conditions plainly: the firmware is old and rarely updated, and the panels are reachable from the public internet.
So the sequence is not an attack in any interesting sense. Somebody scanned for the controllers, found the ones that would talk to a stranger, and changed a number. The number happened to be the temperature of the food supply for a city.
Why it happened twice
France was hit in October. Six weeks later the same class of system fell over in four more countries. In software that gap is enough to patch a fleet twice. Here it was not enough to patch anything, because a meaningful share of these controllers cannot be secured from a keyboard. They need firmware that the vendor may no longer ship, a network that was never segmented, and in some cases a physical modification to the unit itself.
Which means the October wave was not a warning that went unheeded. It was a warning that arrived with a remediation plan measured in truck rolls. Every store, every unit, a technician on site. Nobody completes that in six weeks, and the attackers knew the window was open.
You are about to do this on purpose
Here is why this belongs in a playbook about building with agents rather than in a security newsletter.
The supermarket fridges were exposed by accident, over a decade, by people connecting equipment to a network so they could read a temperature from the office. Nobody decided to put the cold chain on the open internet. It happened one convenient decision at a time.
Right now, deliberately and quickly, companies are doing a more ambitious version of that. An MCP server is a deliberate hole in the wall around an internal system, opened so an agent can reach through it. That is the entire point and it is usually the right call. But the fridges are what the pattern looks like after ten years of nobody revisiting it, and the agent version compounds faster, because an agent probes every endpoint you give it rather than the three a human remembers.
The failure mode is not that your agent turns malicious. It is that you connected something you had forgotten was reachable, and it did exactly what it was asked. The same inventory problem is why 24,664 generated apps hand over database rows to anyone who loads the page. Both frontier labs have already published incidents where agents reached further than intended, and in neither case was the model the surprising part. The surprise was the inventory.
What to do about it this week
Scan your own perimeter from outside it. Not an audit, not a questionnaire. Point a scanner at your public address space from a machine that is not on your network and read what answers. The supermarkets did not know their panels were listening. Most companies do not either.
Write down what an agent can reach, not what it is allowed to reach. Those are different documents and only one of them is true. Permissions describe intent. Reachability describes the blast radius. If you only have the first, you have a policy and not an inventory.
Assume the remediation is slower than the exposure. The lesson of the six-week gap is that discovery speed and fix speed are not related. When you find something exposed, the honest next question is how long a fix physically takes, and whether you can cut the connection in the meantime rather than waiting to secure it properly.
Segment before you connect, not after. Every one of these controllers would have been a non-event behind a network boundary. This is cheap at design time and close to impossible to retrofit across hundreds of sites, which is precisely the position the supermarkets are in now.
A refrigerator does not need to reach the internet and an internal tool does not need to reach every agent. Both get connected anyway, because somebody wanted a number on a dashboard and the quick way worked.
Go find your fridges.


