Between November 2 and 3, 2026, refrigerated display cases in dozens of supermarkets across nine European countries malfunctioned, soaring to 50 degrees Celsius, and just days later, on November 5, OpenAI unveiled GPT-6 Cyber, an artificial intelligence model trained to hunt for vulnerabilities and produce working exploits (programs that take advantage of a flaw to penetrate a system).

The estimated average time to exploit a vulnerability, before or after its patch, according to Mandiant.
At first glance, the two events seem to belong to completely separate worlds: on one side, meat counters and tubs of frozen food thrown out because of compromised industrial controllers; on the other, OpenAI's new model for cybersecurity.
Yet the coincidence raises a question: is there a link between the raid on refrigeration systems and the rise of these new AI models? As far as is known, nothing indicates that artificial intelligence was used in the supermarket attacks. The real common ground is speed. The time between the disclosure of a security flaw and its exploitation in the wild has collapsed, and AI is compressing that window even further.
Contents
- The 50-degree refrigerators: what happened and where AI fits in
- Then and now: the collapse of time-to-exploit
- Claroty's test: twenty-eight minutes and six dollars of AI
- The defensive paradox: GPT-6 Cyber and OT infrastructure
- What to do immediately: isolation, credentials, and lifecycle management
- How we handle security: QTNOVA's approach
- Speed is the new security metric
The 50-degree refrigerators: what happened and where AI fits in
In the early hours of Monday, November 2, 2026, dozens of supermarkets across Europe—from Belgium and Sweden to Italy, where cases were recorded in stores in Sicily, alongside intrusions affecting a catering company in Tuscany and a frozen-food producer in Emilia-Romagna—discovered their refrigerated displays turned into ovens. It was not a particularly sophisticated attack: according to George Nott, technology editor at The Grocer, relatively basic security weaknesses were enough for the attackers: login pages open to the Internet, vulnerabilities already patched by manufacturers, and weak passwords. And many systems, Euronews reports, apparently opened with factory default passwords[1].
In essence, criminals tracked down the network addresses of the digital controllers that remotely manage industrial refrigeration systems. Once inside remote control panels that lacked proper protection, they remotely triggered defrost mode or altered set temperatures (defrosting melts ice: in the French attacks in October, reports indicate it was extended in some cases from one hour up to eight hours), driving temperatures up and causing stored goods to spoil.
So artificial intelligence had nothing to do with it? Not quite. As far as is known, it played no role in the attacks, but it does have to do with the context in which they happen. As Richard Horne, head of the UK's National Cyber Security Centre, pointed out, artificial intelligence did not have to invent new ways to strike us: its power lies in enabling known vulnerabilities that companies had tolerated for years to be discovered and struck at previously unimaginable speed and scale[2].
During those same months, however, AI surfaced in other attacks. Alex Stamos, chief security officer at Cognition (an AI software company) and a lecturer at Stanford, told The Verge's Decoder podcast about an October of back-to-back breaches[3]: some, like the hit on Binance that he attributes to North Korean hackers, were primarily social engineering (deceiving people rather than machines); in others, AI was at the heart of the attack.
CrowdStrike describes one case in detail: between late September and early October 2026, an attacker, likely Chinese, stole data from South Korean financial organizations with ARTEX, an open-source Chinese tool that assigns penetration tests (the attack simulations a company uses to test its defenses) to AI agents, programs that work on their own, and with language models such as DeepSeek and GLM-5.3[4]. Such tools, CrowdStrike concludes, allow an attacker to pull off multiple intrusions in a short amount of time.
Then and now: the collapse of time-to-exploit
To understand the scale of this shift, one must be clear on what an exploit is (the program or sequence of commands that takes advantage of a security flaw to gain unauthorized access or alter a system's operation). An exploit is just one link in the so-called kill chain, a term borrowed from military parlance: the steps an attacker takes to achieve an objective, whether stealing data, shutting down a plant, or siphoning off money. Until now, Stamos explains, research was done ahead of time: you found the flaw, wrote the exploit, and only then used it against one target or thousands. With AI, that chain is reversed: a system surveys targets, finds flaws, and writes exploits in real time.
As long as that was the case, weeks or months passed between the moment a vulnerability was discovered and when someone could craft a reliable exploit. Those defending their networks had a reasonable window of time to test and deploy patches (software updates that fix issues).
That time window has now virtually closed. According to Mandiant's historical analyses, in 2018–2019 the average time required to exploit a vulnerability was 63 days, dropping to 32 days in 2021–2022, before plunging sharply in 2023 to an average of just 5 days[5]. In M-Trends 2026 findings, the estimated average time before exploitation even reached -7 days[6]: this means attackers routinely exploit zero-day vulnerabilities (flaws unknown to the software vendor or lacking a fix) before the vendor can even release a patch. For this reason, Stamos notes, the premise that security has relied on for twenty years—that it is enough to receive a patch and apply it—no longer holds: exploitation happens too quickly.
In parallel, the volume of reported flaws continues to climb. Google's Threat Intelligence Group found that vulnerability disclosures doubled over the course of 2026, surpassing 10,000 monthly reports between July and August. Nor is it just new flaws: according to Google researchers, attackers may be using language models to compare software versions and newly released patches, weaponizing already-patched flaws within days against those who postpone updates.
Projections for the coming months confirm this trajectory. Google's Threat Intelligence Group counted 141 exploited vulnerabilities between January and August 2026, already surpassing the total for all of 2025 (127), and forecasts that AI-assisted vulnerability discovery and exploitation will continue to grow in the short to medium term[7]. In the UK, the NCSC anticipates an unprecedented surge in cyber incidents over the coming months.
Claroty's test: twenty-eight minutes and six dollars of AI
The clearest demonstration of this economic and temporal imbalance comes from an experiment conducted by security researchers at Claroty Team82 on a Danfoss controller from the same family targeted in European supermarkets. Researchers tasked an AI-based software agent with reproducing three known vulnerabilities on the platform, starting from publicly downloadable firmware packages (the embedded software that runs the hardware), formal bug descriptions, and analysis tools.

The stages of an attack using Danfoss controllers as an example: the vulnerabilities are those documented by Claroty; the exact path of the real-world attacks is unconfirmed.
The result: in approximately 28 minutes and for about $6 in AI compute, the agent reconstructed the attack chain and executed commands with root privileges (the highest level of operating-system control, allowing any configuration to be rewritten)[8].
To be completely candid, this was a laboratory test conducted on already-known flaws under the supervision of a human expert, who had to step in when the agent took a wrong turn to steer it back on track. The researcher remained essential. But the test shows just how little time and money someone with moderate skill needs in 2026 to reproduce a known flaw: as Claroty concludes, low costs, minimal time, limited human intervention, and internet-exposed devices make reproducing known vulnerabilities increasingly practical for attackers.
And who foots the bill? Often companies nobody has ever heard of. According to Stamos, criminal groups using ransomware (programs that lock up a company's data and demand a ransom) have for years mainly targeted little-known but important mid-sized companies: when the largest American distributor of medical and dental supplies was hit, dental practices across the country canceled appointments for lack of gloves. Just as with the supermarkets, the damage falls on those downstream.
The defensive paradox: GPT-6 Cyber and OT infrastructure
Right in the middle of this landscape comes the announcement of GPT-6 Cyber. According to OpenAI, the model scored 48.5% on the ExploitGym benchmark (up from 32.1% for its predecessor) and was able to identify 81 previously unknown severe flaws in the Firefox browser, as well as generate working exploits for 948 vulnerabilities in the Linux kernel, at a compute cost of $1,279 per exploit; for now, access is restricted to a select group of Daybreak Red customers, with a new subgroup for critical infrastructure defenders[9].
Yet while the most prominent models, such as GPT-6 Cyber, come with all the expected restrictions, Chinese open models (open-weight: downloaded and run on your own computers, beyond the control of those who created them) can be used by anyone. One example is GLM-5.3 from China's Z.ai: according to CAISI, the artificial intelligence center at the US NIST, it is the strongest open model in cyber ever released, but it lags about four months behind the leading US models; and in Anthropic's tests, it built complete exploits almost as often as Claude Mythos Preview (50 successful attempts out of 410, versus 56), the model Anthropic had reserved for selected defenders[10].
Its safeguards, Anthropic also found, can be bypassed with simple techniques between 64% and 100% of the time, and within days of its release versions stripped of them were already circulating. It is no coincidence that GLM-5.3 was among the models used in the ARTEX attack.
Whether a model is restricted or open, its ability to analyze code poses a massive dilemma for the world of operational technology, known as OT (short for Operational Technology, the hardware and software dedicated to controlling industrial machinery, heat exchangers, valves, and cold-storage units). An IT server is updated regularly; but, as Gianni Cuozzo, CEO of Exein, put it, the greatest worry is “connected devices that remain in the field for years and can be controlled remotely.”
Data from the October 2026 Clusit Report, based on a sample of 75 organizations that completed an OT security assessment, clearly shows this gap: while 80% report using a VPN (a virtual private network that encrypts and secures remote connections) for remote OT access, only 21% apply validated software and operating-system updates to OT/SCADA devices (systems that control and monitor physical processes), and just 25% enforce an equipment lifecycle policy[11]. When AI models accelerate the reproduction of known flaws, leaving non-upgradable equipment online is the equivalent of leaving a door wide open.
What to do immediately: isolation, credentials, and lifecycle management
Protecting yourself does not require magic formulas, but rather rigorous cyber hygiene that treats physical systems as an integral part of the enterprise attack surface. The immediate countermeasures come down to four operational priorities:

Illustrative example: the same controller, exposed and protected, against identical access attempts.
- Shut down Internet-exposed interfaces. No industrial control panel, remote management unit, or administrative portal should ever be accessible simply by typing an IP address into a web browser. All external access must go through a VPN protected by multi-factor authentication.
- Eliminate default credentials and segment networks. Factory default passwords apparently granted entry to many of the affected refrigeration systems. In addition, the network connecting physical machinery must be segregated from the corporate office network, ensuring that a compromised workstation cannot reach the production line.
- Audit and replace end-of-life systems. Keeping hardware or software that is no longer supported by the vendor in operation is an unsustainable risk. If a controller or software release no longer receives security patches, the only right choice is to schedule its replacement or isolate it completely from the outside network.
- Find your own flaws before others do. Using AI-driven scanning tools and running tests that simulate attacker behavior makes it possible to uncover weaknesses before they are hit. GPT-6 Cyber was built for precisely this purpose (finding vulnerabilities and proposing fixes), and the NCSC emphasizes how AI enables flaws to be identified and resolved faster than ever. For a company, this translates to periodically scheduling vulnerability assessment or penetration testing activities. It is the same principle followed by Google, which maintains an internal Red Team: hackers tasked with attacking the company itself by simulating various adversaries, from nation-states to individual criminals and rogue employees, to locate weak spots before real attackers do[12].
How we handle security: QTNOVA's approach
At QTNOVA, we build and manage Odoo environments for our clients with a clear principle: we rely on Odoo, a modern and open platform, and directly oversee its entire lifecycle through structured infrastructure automation processes.
- Continuous updates with zero downtime. Enterprise management software cannot be left to age: we keep Odoo constantly up to date to get the latest security fixes right away. For clients supported under our AMS maintenance agreement, applying security patches is included in the service and operational downtime is zero.
- Hardened, auditable access. The administration of our servers is not reachable from the Internet. Access is permitted solely from within our encrypted VPN, where every permission requires a tracked, reviewed change in configuration code, and private cryptographic keys never leave individual team members' workstations.
- Backup and recovery. Security is also measured by the ability to get back up and running after an incident. We maintain two independent backup pipelines stored in a geo-redundant space, and recovery has its own procedure, to be tested in a separate environment that does not touch production.
- Ahead of deadlines. We upgrade our servers' platform before its versions reach their deprecation date: that way, in the event of a disaster, we can still rebuild it from scratch.
Speed is the new security metric
The attacks on refrigeration systems and the debut of GPT-6 Cyber tell the same story from two different angles: in a lab, an AI agent reproduced an exploit for a known flaw in under half an hour, while many devices remain deployed in the field for years.
If your ERP runs on a version that no longer receives security updates, the risk is no longer theoretical. Want to know where your Odoo deployment stands? Let's talk.
Want to know more? Leave us your email below and we'll get back to you as soon as possible.
Sources:
- Euronews Next — “Hacker sondano i sistemi di refrigerazione europei prima degli attacchi a supermercati e banche del sangue” (November 7, 2026) – it.euronews.com
- NCSC — “Richard Horne's speech at the NCSC Annual Review 2026 launch” (November 3, 2026) – ncsc.gov.uk
- The Verge — “AI broke the cybersecurity world. What’s next?” (November 7, 2026) – theverge.com
- CrowdStrike — “Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance” (October 7, 2026) – crowdstrike.com
- Mandiant — “How Low Can You Go? An Analysis of 2023 Time-to-Exploit Trends” (October 15, 2024) – cloud.google.com
- Mandiant — “M-Trends 2026: Data, Insights, and Strategies From the Frontlines” (March 23, 2026) – cloud.google.com
- The Record — “Google: Vulnerability disclosures double to 10,000 per month as AI fuels exploitation” (September 30, 2026) – therecord.media
- Claroty Team82 — “Firmware to Root in 30 Minutes: How AI is Changing Threat Intelligence” (October 13, 2026) – claroty.com
- The Decoder — “OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them” (November 5, 2026) – the-decoder.com
- Anthropic — “GLM-5.3 and the spread of advanced cyber capabilities” (September 29, 2026) – anthropic.com
- Cybersecurity360 — “Rapporto Clusit 2026: dati e rischi cyber per le imprese italiane” (October 29, 2026) – cybersecurity360.it
- Google — “Google's AI Red Team: the ethical hackers making AI safer” (July 19, 2023) – blog.google