Anthropic research finds serverless smart worms replicating models on home devices

Smart TVs and home refrigerators could become, within months, a self-replicating environment for miniature models, without needing to connect to large data centers.This direct warning did not come from an external monitoring body, but was documented by researchers at Anthropic in a recent report that revealed small smart models are approaching sufficient capability to copy themselves across Internet-of-Things devices, opening a unprecedented door for malicious software to operate with full autonomy.

The Red Team for advanced security testing at Anthropic conducted practical experiments involving American and Chinese models, and researchers William Zou and Daniel Freeman concluded that AI worms could find a primary target in home devices and small-store networks, given their constant internet connection and historically lacking the simplest security hardening standards. Although the current experiments have not recorded a model capable of completing a reliable replication cycle on a smart screen due to its small size, the rapid development pace suggests a possibility that this capability could appear in practice before the end of 2026.

The technical danger lies in the fundamental difference between conventional malware and that driven by intelligent agents; traditional worms require a human programmer to modify their path and direct attacks, whereas an AI agent can analyze unfamiliar device environments, adjust its behavior when the first attempt fails, and continue spreading automatically without a human operator, a point confirmed by AI security specialist Roman Yambolski of Loyola University, who warned of the difficulty of stopping these systems due to the absence of a central server or single controlling entity.

The shift toward the spread of lightweight agents beyond direct human oversight imposes a new defensive equation in Middle-East markets, particularly in Gulf cities that rely on massive digital density in smart homes and interconnected commercial facilities. This scenario means that corporate cybersecurity strategies can no longer rely solely on protecting cloud gateways and regional data centers, but are now required to reassess the auditing of edge-connected devices and to isolate Internet-of-Things networks from sensitive operational infrastructure to reduce potential breach surfaces.

The race is inseparable from the battle to protect intellectual propertySimultaneously with the disclosure of self-replication vulnerabilities, Anthropic announced a reduction in unauthorized data-extraction operations by between 90 and 99 percent, referring to protective measures against extraction attacks that it accused Chinese competitors of using to clone its models' capabilities, confirming that the struggle for model control no longer takes place solely within supercomputing labs, but is rapidly extending to devices distributed in every corner of daily life.