Anthropic Warns Self-Replicating AI Could Target Smart Home Devices
Anthropic researchers warn that a new class of self-replicating, AI-powered malware could eventually move through smart-home and other connected devices. Unlike conventional computer worms, which follow fixed instructions and rely on known vulnerabilities, an “AI worm” could adapt to different operating systems and hardware limits, find ways around security updates and learn how to function in unfamiliar environments.
In a controlled experiment, Anthropic distributed an AI model across 15 inexpensive Android TV boxes, showing that some AI workloads do not require a large data center. The model then infiltrated simulated Android devices and replicated across all eight targets in four of five trials; in the fifth, it reached seven. Researchers estimate such software could take one to two days to establish itself on a previously unknown device. A growing network of compromised devices could form a decentralized botnet, making the malware harder to stop by shutting down a central server. Anthropic says smaller models may become capable of spreading in real-world conditions later this year, although its tests deliberately used weak safeguards.
The risk is relevant as connected systems proliferate: about half of German households use smart-home applications. In France, attackers recently manipulated supermarket refrigeration controls, pushing temperatures as high as 50 degrees Celsius and causing spoiled food, fires and technical damage. CERT-FR confirmed that remote-control systems from Danfoss and other manufacturers were affected, possibly through outdated security software. There is no evidence that AI malware was involved, but the incident demonstrates the physical consequences of insecure connected equipment. French authorities urged regular updates, essential security patches and disabling unnecessary internet access.
Continue in the app — vote & join in ➔