Friends ask me some version of this a lot: how would a bad actor with super powerful AI actually hurt anyone in the real world?
It usually comes with some skepticism. The All-In Podcast guys have talked about this and tend to wave it off as a straw man, like everyone concerned is picturing the Terminator.
So here is a much more mundane example from this week.
What happened
The New York Times reports that a cyberattack hit the refrigeration systems of more than 400 supermarkets in France. That is about 300 Intermarché and Netto shops, 80 from Coopérative U, and 30 from Monoprix.
The attackers went after the software that lets refrigeration companies control the fridges remotely. Some shops had to throw out all of their fresh and frozen food. The head of Coopérative U said a failed system can cost a small supermarket tens of thousands of dollars, and a hypermarket roughly $220,000.
Nobody has claimed responsibility. Danfoss, one of the companies that makes these systems, said the affected ones were often outdated and no longer received security updates.
This is definitely not the first time
Software has been breaking physical things for a while now:
- Stuxnet, discovered in 2010, destroyed roughly a thousand uranium centrifuges in Iran.
- In December 2015, attackers cut the power to about 230,000 people in Ukraine.
- In 2021, ransomware hit the business systems at Colonial Pipeline. The company shut the pipeline down, and gas stations across the US Southeast ran dry.
Why this one sticks with me
Stuxnet took a nation-state budget and years of work. The freezers took an old controller that was reachable from the internet.
To be clear, nothing reported so far says AI had any part in this attack. It didn't need one, and that is why it is a useful example.
An attack like this is mostly patient, tedious work. Someone has to find the exposed devices, read through old firmware, and work out which setting does the damage. Today the limit on attacks like this is how many skilled people want to do it, and how much time they have. Patient, tedious work across a long list of targets is exactly what AI models are getting good at.
And the world is full of freezers. Building HVAC, water pumps, parking gates, the fridge at your pharmacy. Nobody thinks of these as targets, so they run old software with remote access switched on.
So when a friend asks what a bad actor with powerful AI could do, my answer has no robots in it. It looks a lot like this week in France, at more places, more often, and from people who couldn't have pulled it off on their own before.
What is the most boring internet-connected thing at your work that would cause real trouble if someone turned it off?
Top comments (8)
urg yes... true story.. but whyy
Why not? Though I don't know why either lol
Diabolical.
"An attack like this is mostly patient, tedious work." This observation will stick with me. To have the idea, work on it, iterate on it and then deploy it, to me, the effort on it blows my mind.
Yeah, I agree with your point on the old controllers with remote access on. But the most boring thing I have on the internet is my payment webhook endpoint. If it goes down I don’t realize it until hours later and orders are unpaid and they are for money already paid. If a forged call was accepted, multiple orders would be marked paid without any payment and there’s no fixing that. Many teams don’t even bother checking the signature on their endpoint because that’s a minor problem.
AI agents will have the same problem. Everyone wants more access to computers and servers and accounts with AI. What I’m more interested in is how you can reliably prevent these agents from doing something stupid. I’m much more interested in the security than the demos of AI agents.
Good point about it not being sci-fi — these are mundane systems that suddenly become critical. The pattern repeats: legacy devices out of patch support, remote access exposed, and the blast radius is measured in spoiled food and real money. It’s the same class of failure as “unmaintained dependencies”, but with fridges instead of npm packages. The scary part is how many of these industrial controllers are still internet-reachable with default creds or no auth. The story reinforces one rule: if something can spoil, leak, or stop production when it goes wrong, it needs to be on an air-gapped network or behind a very narrow authenticated gateway.
The shared credential detail matters more than the outdated software framing: if the maintenance password was the same across all 400 stores, the attack didn't need to find 400 vulnerabilities — it needed to guess one date. And a date-as-password means the maintenance account was designed for convenience, not for audit: no unique credential per store, no way to tell which technician accessed which store when, no sealed log of what changed. The evidence layer isn't missing because the system is old — it's missing because nobody designed it to know who did what. A unique credential per store, sealed at provision time, would have made the blast radius visible as 400 separate access events instead of one shared password. The receipts don't prevent the attack; they make the blast radius auditable after the fact — which is the difference between "we think it was one password" and "here are the 400 stores, the 400 access times, and the one credential they shared."
Chased the count, since the wires disagree with each other on the same day. The NYT breakdown you cite (300 Intermarché/Netto, 80 Coopérative U, 30 Monoprix) doesn't match Le Monde's, which puts the 80 at Carrefour, about 100 at Coopérative U, several E.Leclerc stores on top of that, and around 30 Monoprix. And AFP's count for Intermarché is 150-200, not 300 (BBC carries it). So 'more than 400' is the safe part; the per-chain numbers are each chain's own statement, and nobody has published the full list yet. One detail that makes your point sharper: Le Monde reports the attackers stretched the defrost cycle from one hour to eight, which pushed the alarms later. The tedium was the weapon.
Making attacks like this cheaper and easier to repeat is scary enough without the Terminator.