Paul Roebuck.co.uk

Locks and Keys

Observation 002 — The fridges were on the street

21 November 2026 · 7 min read

Cover card: Locks and Keys, 002. The fridges were on the street. A fridge standing on the street with a gold key left in its door. Paul Roebuck.

Locks and Keys · break-ins and break-outs, observed weekly from The Observatory · written Friday 13 November 2026 · Paul Roebuck

The position. AI doesn’t break the law. People do. AI tries doors. Where a key is lying about, it goes through. This week’s biggest break-in story had no AI in it at all, as far as anyone can show. I include it because it is exactly the kind of door AI agents find at machine speed. And because headlines keep bringing AI into it anyway.


1. Seven hundred shops, and the door was on the street

Leg 1 · Break-in · the door AI agents find. No source says AI was used.

OBSERVED. Overnight on Sunday 11 into Monday 12 October, someone got into the controllers that run supermarket fridges and freezers in France. They changed the defrost settings to high temperatures. Food spoiled. There were at least 15 fires. France’s cyber crisis centre (the C4, which includes the national cyber agency ANSSI) puts it at about 700 sites, mostly supermarkets: Casino, Leclerc, Coopérative U, Metro, Carrefour and Intermarché among them. Butchers, an abattoir and food factories were hit too.

The first counts were lower. On 13 October Al Jazeera carried the retailers’ own figures: at least 300 Intermarché and Netto shops, more than 100 Coopérative U, 80 Carrefour, about 5% of Monoprix. That same day, by AP’s account, the finance minister told Parliament about 700.

The kit named first was Danfoss refrigeration controllers. France’s computer emergency team, CERT-FR, issued an alert on 12 October naming Danfoss’s AK-SM and AK-SC ranges, later adding Copeland and Dixell systems. It said exposing a management screen to the internet goes against good practice. It named no specific flaw. As reported by SecurityWeek, Danfoss said most of the units hit were old: the AK-SM 800, out of support since May 2021, and the AK-SC 355, discontinued in 2016. Some current AK-SM 800A units on outdated firmware were hit as well.

The doors had been counted in advance. In August the security firm Claroty published flaws in the AK-SM 800A, fixed in new firmware, and, citing a scan by Censys, put 2,765 of those controllers on the open internet. On the Monday of the attack a researcher, Baptiste Robert of Predicta Lab, counted 2,752 Danfoss systems exposed online, 1,590 of them in France. Nobody has published exactly how the attackers got in.

On 9 November France named who turned the key. The C4 attributes the attack to Shahid Kaveh, a cyber unit of Iran’s Revolutionary Guards. Its evidence: the same attack servers were scouting Schneider Electric industrial kit; the attackers used FOFA, a search engine for internet-connected devices; an outlet tied to the unit publicised the attack on 13 October; and there were traces of Farsi. The report says the unit has had scripts since at least June 2024 to scan for, get into and control industrial equipment. This is France’s attribution. I have not seen its evidence beyond what the report lists.

On AI, every source I read says the same thing: nothing. The CERT-FR alert, the C4 report, Claroty and Danfoss do not say AI was used. The head of the UK’s National Cyber Security Centre, Richard Horne, said it was not an AI attack. The French systems, he said, were old, no longer updated, connected straight to the internet, and still on default passwords (as reported by AP, 3 November; AP gives his words in its own paraphrase, not as a quotation).

READ. A person chose the target. A search engine listed the doors. The unit, France says, already had scripts for walking through doors like these. The locks were years out of date, and, by the UK cyber chief’s account, some still had the factory key in them. As far as the sources show, no AI was needed. I file it here because this is precisely the door an AI agent finds at machine speed: a login screen on the open internet, on a box its maker stopped supporting years ago. Take AI out and the door was still open. Put AI in and it opens faster.

OBSERVED, the aftermath. A second wave, first reported on Monday 2 November, hit more than 115 stores in six European countries (Euronews, 3 November; BleepingComputer, 9 November). France attributes only the October attack to Iran. On 5 November the US cyber agency CISA, the FBI and three other US departments issued a joint fact sheet. Its advice: take refrigeration controllers off the public internet, change default passwords, patch, and retire kit the maker no longer supports.

2. The locksmith used AI, and held the details back

Leg 1 · Break-in, defender’s side · AI found the door.

OBSERVED. In late October Fortinet, which makes firewalls and network security kit, shipped fixes for FortiOS, FortiManager, FortiAnalyzer and others. Its security chief, Carl Windsor, wrote that the fixes came from testing with frontier AI models; Help Net Security reported that people then checked the findings. Fortinet held the details back until 10 November, to give customers time to patch before the details went public. On 10 November it published the advisories: more than 60 that day in its own listing. Nearly all credit Fortinet’s own team, working "with the use of the frontier AI model: Claude Mythos", which is Anthropic’s. Several are rated critical, and at least six of the critical ones need no login. None that I read is marked as exploited. Fortinet’s blog warns that AI can link several minor flaws into one serious one, and can help attackers turn a published patch into a working attack in hours or days.

READ. This is the good hand on the same drive. The machine found the doors for the locksmith first. But the clock is the point. Once a fix is public, it shows where the hole was. If you haven’t patched, you are now the door.

3. Nearly a thousand doors in Linux, found by agents

Leg 1 · Break-in, defender’s side · AI found the door.

OBSERVED. On 4 November OpenAI published a report on a pipeline of AI agents run in October against the Linux kernel, the core of the Linux operating system. The agents produced 948 working exploits, each turning an ordinary user on the machine into its administrator, at a cost of about $1.2 million in computing. People reviewed each exploit and fix before sending them to the Linux maintainers. Ten days in, 61 were already fixed. Each of these needs a foothold on the machine first. CSO Online reported the same figures on 5 November.

READ. The same drive, pointed at the locks by the people who make them. Nearly a thousand fewer doors, if the fixes land. And a reminder that whoever runs this next may not send the fixes to anyone.

Leg 2 · Break-out — none verified this week.

4. Lying about keys: "could AI make it worse?"

Leg 3 · Lying about keys. A soft case: the headlines bring AI in; neither says AI did it.

The framing. On 3 November Euronews ran: "Why are supermarket freezers being hacked across Europe and could AI make it worse?" Cybernews ran the headline "Hackers could disable your local supermarket fridge for just $6", its web address naming Claude, Anthropic’s AI.

OBSERVED. Euronews does not say AI was used in the attacks. It says many of the controllers hit in France were older models that no longer get security updates, despite flaws known since 2025. The $6 comes from Claroty, published on 13 October. In its lab, a researcher gave Claude public details of three flaws that were already known and already fixed, plus firmware from Danfoss’s own website. In about 28 minutes, for about $6 of AI time, it rebuilt a working attack. It got stuck on the login bypass and a person had to put it back on track. Claroty says the route used in France "remains unconfirmed". The UK’s cyber chief said it was not an AI attack.

READ. The headlines lean on the machine. The sources hang it on the fridges: on the internet, out of support, factory passwords in place, and, by France’s account, a state unit with a device search engine. AI could have made it faster. It didn’t need to. Blame the machine and you stop looking at who left the key in the door.


Why it does it: my postulate

POSTULATE. Nothing new this week, and that is the point. The door finder doesn’t care whose hand is on it. By France’s account, a state unit with a search engine got through about 700 doors, and nobody has shown it used AI. An AI in a lab, with a person putting it back on track, rebuilt a working attack on the same make of controller for the price of a sandwich. Firms are now using the same machines to find their own doors first. The drive is the same. What changes is who points it, and whether anyone locked up before it arrived.

The lock to check today

If you run anything with a screen that controls something physical (fridges, boilers, pumps, doors), ask one question: can it be reached from the open internet? If yes, take it off, or put it behind a proper VPN. Change the factory password. If the maker no longer updates it, plan to replace it. And if you run Fortinet kit, the details of last month’s fixes went public on 10 November. Patch today.

Typed by Bob, my AI. Conceived, checked and published by the human in the loop.


Sources

  1. CERT-FR, Incident de sécurité ciblant les contrôleurs à distance des installations de réfrigération, CERTFR-2026-ALE-012, first version 12 Oct 2026, last version 16 Oct 2026 — https://cert.cyber.gouv.fr/alerte/CERTFR-2026-ALE-012/ (read at source)
  2. C4 / CERT-FR, Activités cyber offensives menées par des acteurs iraniens liés au CGRI, CERTFR-2026-CTI-009, 9 Nov 2026 — https://cert.cyber.gouv.fr/cti/CERTFR-2026-CTI-009/ and the full report https://cert.cyber.gouv.fr/uploads/CERTFR-2026-CTI-009.pdf (read in full at source)
  3. Al Jazeera, Hundreds of supermarkets in France hit by ’unprecedented’ cyberattack, 13 Oct 2026 — https://www.aljazeera.com/news/2026/10/13/unprecedented-cyberattack-hits-french-supermarkets-causing-food-losses (read at source; the retailers’ figures, as given to AFP)
  4. BleepingComputer, Cyberattack shuts down refrigerators at French supermarkets, 12 Oct 2026 — https://bleepingcomputer.com/news/security/cyberattack-shuts-down-refrigerators-at-french-supermarkets (read at source; Baptiste Robert’s exposure count) · BleepingComputer, France blames Iranian hackers for supermarket attacks that sparked fires, 9 Nov 2026 — https://bleepingcomputer.com/news/security/france-blames-iranian-hackers-for-supermarket-attacks-that-sparked-fires (read at source; the second wave, and France’s attribution covering only October)
  5. SecurityWeek, Supermarket refrigeration attack widens to Copeland, Dixell controllers, 15 Oct 2026 — https://www.securityweek.com/supermarket-refrigeration-attack-widens-to-copeland-dixell-controllers/ (read at source; Danfoss’s statement on the models is reported here; not found on Danfoss’s own site)
  6. ComplexDiscovery, 14 Oct 2026 — https://complexdiscovery.com/refrigeration-hack-affected-about-700-french-supermarkets-government-says-as-danfoss-interfaces-stay-online/ (read at source; reports AP’s account of the finance minister’s figure to Parliament, 13 Oct, and Danfoss’s words to Reuters and franceinfo)
  7. Claroty Team82, Freeze the controller, defrost the food, 9 Aug 2026 — https://claroty.com/team82/research/freeze-the-controller-defrost-the-food-uncovering-vulnerabilities-in-danfoss-refrigeration-controllers · Claroty Team82, Firmware to Root in 30 Minutes: How AI is Changing Threat Intelligence, 13 Oct 2026 — https://claroty.com/team82/research/firmware-to-root-in-30-minutes-how-ai-is-changing-threat-intelligence (both read at source; Claroty sells security for industrial systems)
  8. CISA, FBI, HHS, USDA and DoT, US Government Partners Urge Hardening Industrial Refrigeration and Cold Chain Monitoring Systems, fact sheet, 5 Nov 2026 — https://www.ic3.gov/CSA/2026/261105.pdf (read in full at source)
  9. AP, UK will face AI cyberattacks on critical infrastructure by 2028, official says, 3 Nov 2026 — https://mynews13.com/fl/orlando/ap-top-news/2026/11/03/uk-cyber-chief-says-the-uk-will-face-ai-cyberattacks-on-critical-infrastructure-by-2028 (read at source, as syndicated; Horne’s words on France are AP’s paraphrase)
  10. Euronews, Why are supermarket freezers being hacked across Europe and could AI make it worse?, 3 Nov 2026 — https://www.euronews.com/2026/11/03/why-are-supermarket-freezers-being-hacked-across-europe-and-could-ai-make-it-worse (read at source) · Cybernews, Hackers could disable your local supermarket fridge for just $6 — https://cybernews.com/ai-news/hackers-disable-supermarket-fridge-claude-ai-costs-6-dollars/ (headline and web address only; the page refused to load; date not seen)
  11. Fortinet, Carl Windsor, Fortinet releases security hardening release for multiple products, 21 Oct 2026, updated 10 Nov 2026 — https://www.fortinet.com/blog/ciso-collective/fortinet-releases-security-hardening-release-for-multiple-products · Fortinet PSIRT advisories published 10 Nov 2026, and their credit lines — https://fortiguard.fortinet.com/psirt (both read at source) · Help Net Security, 26 Oct 2026 — https://helpnetsecurity.com/2026/10/26/fortinet-patches-critical-fortios-flaws-found-with-ai-help (read at source; "validated by humans" is Help Net Security’s wording) · Anthropic, Claude Mythos — https://anthropic.com/claude/mythos (read at source; the model is Anthropic’s)
  12. OpenAI, A thousand fewer paths to root, 4 Nov 2026 — https://openai.com/index/a-thousand-fewer-paths-to-root/ (read at source) · CSO Online, OpenAI built an AI pipeline to find, exploit and patch Linux flaws, 5 Nov 2026 — https://www.csoonline.com/article/4245265/openai-built-an-ai-pipeline-to-find-exploit-and-patch-linux-flaws.html (read at source)
  13. Also read: OCCRP, Hackers Turn French Supermarket Refrigerators Into Furnaces — https://www.occrp.org/en/news/hackers-turn-french-supermarket-refrigerators-into-furnaces · ITPro, French fridges hacked in country-wide supermarket attack, 13 Oct 2026 — https://www.itpro.com/security/cyber-attacks/french-fridges-hacked-in-country-wide-supermarket-attack (both read at source; neither says AI was used)

Comments

Comments appear straight away and anyone can read them. Please don’t share anything you’d want kept private.

  1. Loading comments…