From freezers to a hosting provider: what a month of strange cyberattacks actually shows

Between October 9 and November 9, 2026, attackers raised the temperature in supermarket freezers across Europe, wiped a hosting provider's servers, and abused a vendor's cloud app to try to reach thousands of Microsoft 365 tenants. These incidents have little in common on the surface. This piece connects them through what is documented and keeps separate what is not.

Freezers: two waves in three weeks

In the night of October 11–12, attackers set refrigeration controllers in French supermarkets to defrost mode. Palo Alto Networks Unit 42 writes that this left "hundreds of supermarkets across the region having to throw away spoiled food" and, citing multiple sources, that temperatures in targeted units rose "as high as 50°C (122°F)" (Unit 42). The count of affected stores varies by source: Unit 42 uses about 700, while other outlets report lower figures.

A second wave hit overnight on November 1–2 and reached at least ten more European countries, with 100 to 115 retail locations affected. Euronews lists Belgium, Sweden, the Netherlands, Estonia, Luxembourg, Ireland and Finland, and says the Belgian federal prosecutor opened an investigation.

What is established, and what is not

• Established: the mechanism. Controllers were switched to defrost, and the temperature and cycle length changed. Belgian installer Sabcobel says the November wave used the same mechanism as the October one (via Euronews).
• Not established: how the attackers got in. Unit 42 says speculation points to vulnerable software or default credentials, and calls it unconfirmed.
• Not verified: claims of responsibility. NoName057(16) and Z-Pentest claimed the attacks on Telegram on November 3, and Unit 42 states these claims "remain unverified."
• Danfoss' position: the company, whose software runs on the affected systems, blames customers who did not install updates and says the flaw stems mainly from outdated software (paraphrased by Euronews). Unit 42 does not name vendors or CVEs in its write-up.

Two Danfoss AK-SM controller flaws are in the public record: CVE-2025-41450 and CVE-2025-41451, both rated high severity. Reporting on the attacks mentions them, but we found no source that confirms either was used in these incidents, so we list them as context for operators, not as the cause.

The medical case shows both sides of the problem. The attack came through a supplier's remote-monitoring software, not through the blood service itself. And it was caught because independent monitoring raised alarms and staff moved the plasma, not because anyone could stop the change in the controller.

A hosting provider: Beget

In the night of November 1–2, Russian hosting provider Beget was attacked. Its official status page gives a timeline in Moscow time: management servers reached around 02:00, malware "designed to delete data from the servers" launched at 02:30, all servers disconnected from the network at 02:55. By November 5 the company reported 60% of shared hosting and 40% of Beget-managed dedicated servers restored, and on November 7 the control panel returned with limited functionality (Beget).

Beget has not published the attack vector, the attackers' identity or the final extent of data loss. We found no source linking this attack to AI. The architectural lesson is plain: when one management layer can act on every server, one compromise reaches all customers at once.

Trusted tools: a vendor app and a CLI

• ConnectWise Cloud Backup. Microsoft assesses that the China-aligned actor Storm-0940 was responsible for the ConnectWise supply-chain compromise and took control of identities of multiple ConnectWise cloud apps, then used them against downstream customer tenants. Microsoft counted sign-in attempts to more than 2,250 tenants. That is attempts, not breaches: Microsoft says "the majority of the targeting wasn't successful" and that an affected app or a successful sign-in "does not necessarily indicate that customer data was accessed or exfiltrated" (Microsoft).
• OVHcloud CLI. On October 19 an attacker used a compromised account to replace GitHub release assets for versions v0.11.0 to v0.15.0 of the OVHcloud CLI. The window was about 94 minutes; OVHcloud removed the assets the same day and released a clean v0.16.0. The binary was designed to read secrets from disk, so CERT-EU advises treating any execution in the window as credential exposure (CERT-EU).
• PHP2Shell plus a Linux kernel flaw. Dreamgroup reported on October 26 a chain of PHP2Shell and a kernel bug, CVE-2026-46300, that gives root and lets attackers drop PHP web shells. Dreamgroup attributes compromises of at least 25 PAN-OS firewall or management servers to it (the count is the researchers', not CISA's). CISA added the kernel flaw to its KEV catalog on October 28 (CERT-EU). For the web-interface bugs themselves, see our PHP2Shell guide. A separate PAN-OS authentication bypass, CVE-2026-0257, saw ransomware-linked exploitation attempts reported by Unit 42 on October 20.

Note what PHP2Shell is not: a claim that any PHP site is vulnerable. It concerns specific products that ship PHP components.

Where AI actually appears

The same Microsoft write-up documents something new. Storm-0940 used ARTEX, which Microsoft calls "an open-source penetration testing framework that can use any AI model." It is a multi-agent system that runs a test "from reconnaissance through data exfiltration." The numbers are observations, not victim counts:

Anthropic's September threat report reaches a similar conclusion from another angle. It says AI "has collapsed the labor and tooling gap" between state-backed and individual operators, that humans still set targets and review stolen data, and that none of the operations it describes relied on an entirely novel technique: the change is in the economics (Anthropic).

What we can and cannot claim

• Supported: attackers are using AI agents to run parts of intrusions faster and against more targets. Microsoft and Anthropic both document it.
• Supported: physical systems on the internet are being hit. Two waves of refrigeration attacks in three weeks are documented.
• Not supported: that AI caused the refrigeration attacks or the Beget attack. No source we read says so, and the observed methods (exposed controllers, management-layer access) do not require it.
• Not supported: that attack counts are rising because of AI. A rising count of published CVEs is not a count of successful attacks.

What to do

1. Take controllers off the public internet. Refrigeration, building management and other OT devices belong behind a VPN or an allow-list. Unit 42 found over 5,000 exposed devices from the manufacturers in a CERT-FR advisory.
2. Update, and replace what cannot be updated. Danfoss says outdated software is the main issue; Sabcobel says updates alone will not fix the flaw and hardware changes are needed. Check vendor advisories for the exact models you run.
3. Add independent alarms. The plasma was saved by monitoring and people, not by the controller. A separate sensor that pages a human is cheap.
4. Treat suppliers as part of your attack surface. The EFS case came through a refrigeration supplier's monitoring software. Ask which vendors can change settings in your systems remotely.
5. Keep backups outside the control layer. Backups that the same admin account can delete are one copy, not three.
6. Rotate credentials after a supplier incident, and treat any run of a tampered release as credential exposure.
7. Know what you run. My Stack tracks the products you use and flags them when a CVE moves toward exploitation; the exploitation radar shows what attackers use now.

Data as of 2026-11-09live record →

Sources: Palo Alto Networks Unit 42, Nov 6, 2026, Microsoft Threat Intelligence, Oct 21, 2026, Beget status page, CERT-EU Cyber Brief 26-11, Anthropic, Detecting and countering misuse of AI: September 2026. Also used, as secondary reporting for quotes from EFS and Danfoss: BleepingComputer (Nov 4) and Euronews (Nov 2). Statements of Beget are paraphrased from its English status page.