Anthropic researchers have warned that AI worms could target connected household devices and spread from one device to another. Cyberattacks on French supermarkets illustrate the vulnerabilities of such systems.
Connected electronic devices, part of the "Internet of Things," include items like robot vacuum cleaners, voice assistants, and smart TVs. Recent market research by Bitkom indicates that about half of all households in Germany utilize smart home applications, with this figure on the rise.
Anthropic's report suggests that AI-powered malware could infiltrate connected household devices, replicate itself, and activate on other devices. Traditional computer viruses, known as worms, have existed for decades and exploit security vulnerabilities. However, AI worms would be more complex and potentially harmful, capable of autonomously solving problems and adapting to different devices.
In an experiment, Anthropic researchers distributed an AI model across 15 low-cost Android TV boxes, demonstrating that certain AI applications can run without large data centers. In a controlled test with eight simulated Android devices, the AI model infiltrated one device and replicated itself, successfully reaching all target devices in four out of five attempts.
If equipped with malware, such AI models could adapt quickly to new environments and disrupt operations. They could learn to bypass security measures rather than relying on a single vulnerability. Anthropic developers estimate that an AI worm could establish itself on a new device within one to two days.
A concerning possibility is that AI could link infected devices into a botnet, increasing the speed of spread as more devices are compromised. The decentralized nature of such malware would complicate containment efforts, likened by Anthropic to a digital slime mold.
In a recent incident in France, attackers compromised refrigeration and freezer systems at several supermarkets, causing temperatures to rise to 50 degrees Celsius, leading to food spoilage and technical damage. France's cybersecurity agency, CERT-FR, confirmed that remote-control systems from Danish manufacturer Danfoss were compromised, with investigations ongoing.
Modern supermarkets utilize digital controllers for monitoring refrigeration systems, but some security software was reportedly outdated, which may have been exploited by attackers. The identity and motivation of the attackers remain unclear, and there is no evidence that a self-replicating AI worm was involved in the French attacks. However, Anthropic's research indicates that AI could significantly enhance malware development, posing serious risks to connected devices.
In response to potential threats, France's cybersecurity agency has advised companies to regularly update connected devices, install essential security patches, and disable unnecessary internet access where possible.