The next frontier in AI-driven cyberattacks may be sitting in your kitchen.
A cyberattack targeting refrigeration systems at hundreds of French supermarkets has already caused food losses and reportedly sparked fires. Meanwhile, new research from Anthropic warns that AI models could soon gain the ability to reproduce across everyday internet-connected devices, including smart refrigerators, ovens and televisions, without relying on centralized data centers.
Together, the developments raise a troubling possibility: Cyberattacks that once threatened data and business operations could increasingly manipulate physical equipment, while autonomous AI agents make those attacks harder to detect and contain.
When refrigerators become attack surfaces
According to an October 12 CBS News report, attackers remotely manipulated refrigeration controls across major French supermarket chains, including Intermarché, Cooperative U and E. Leclerc.
At least 300 Intermarché and Netto locations, more than 100 Cooperative U stores and additional retailers were affected. Refrigeration temperatures reportedly climbed after attackers triggered defrosting systems remotely, forcing stores to discard fresh and frozen products.
French prosecutors also confirmed reports of fires associated with the incident.
Investigators identified vulnerabilities involving refrigeration-control equipment supplied by Danfoss. The manufacturer said affected systems included legacy products that no longer receive security updates.
The incident was not publicly established as an AI-powered attack. But it demonstrates the physical consequences that compromised connected equipment can produce, even without artificial intelligence.
Anthropic warns AI could spread through smart appliances
The implications become more concerning alongside research published October 12 by Anthropic.
As CNBC reported October 13, Anthropic researchers believe smaller AI models may be only months away from independently replicating across commercial Internet of Things devices.
In its research findings, Anthropic demonstrated that a 26-billion-parameter model could run across 15 inexpensive Android TV boxes, distributing its computational workload among the devices.
Researchers also tested whether models could discover new systems, establish execution environments and continue reproducing.
Although the experiments did not demonstrate a fully autonomous AI worm operating in the wild, they showed that many necessary capabilities already exist. Anthropic projects that sufficiently capable smaller models could emerge before the end of 2026.
The concern extends beyond televisions. Smart ovens, refrigerators, robotic vacuum cleaners and connected lawnmowers could potentially supply computing resources for distributed AI agents.
Unlike traditional malware, these agents could adapt to unfamiliar hardware, troubleshoot failures and seek alternative routes to replication.
And because their computing resources would be distributed across numerous devices, shutting down a single server would not necessarily stop them.
AI agents are already testing security boundaries
The warning follows other incidents highlighting the growing risks of autonomous systems.
Earlier this month, CYBR.SEC.Media reported suspected AI-assisted attacks against seven South Korean financial institutions that exposed information belonging to more than 67,000 people.
Investigators identified evidence associated with an autonomous penetration-testing tool, although its precise role remains unconfirmed.
At CruiseCon AI/Privacy 2026, former CIA case officer Erin Whitmore warned that attackers could also manipulate information trusted by AI systems, effectively turning legitimate models into unwitting insider threats.
And a September CYBR.SEC.Media analysis highlighted research finding that 25% of organizations had deployed AI agents in production without monitoring.
These examples represent different threat mechanisms, not evidence of one coordinated attack campaign. Their common denominator is growing machine autonomy combined with inadequate visibility and control.
What security practitioners can do to meet this threat
Security teams need to extend established defensive practices to connected appliances, embedded controllers and autonomous agents.
- Inventory connected equipment. Identify smart appliances, refrigeration controllers, IoT gateways and vendor-managed devices. Document firmware versions, support status and external connectivity.
- Segment operational networks. Keep appliances and industrial controllers isolated from corporate endpoints and sensitive systems. Restrict unnecessary east-west communications.
- Eliminate unsupported technology. Prioritize replacement of devices that no longer receive security patches, particularly equipment controlling physical processes.
- Monitor unusual device behavior. Watch for unexpected outbound connections, large model downloads, scanning activity and sustained device-to-device traffic.
- Constrain autonomous agents. Apply least privilege, short-lived credentials and approval requirements for consequential changes. Log agent actions and preserve emergency shutdown capabilities.
- Prepare for physical consequences. Test incident-response procedures involving equipment isolation, manual overrides, temperature monitoring and operational recovery.
Anthropic's researchers specifically recommend detecting anomalous network flows and quarantining infected systems to interrupt propagation.
The French supermarket attacks demonstrate that connected equipment can already become a source of physical disruption. Anthropic's research suggests future attackers may not need to personally orchestrate every stage of compromise.
The refrigerator is no longer just an appliance. It's a connected endpoint with potential physical consequences. And security teams must start treating it that way.
