Artificial intelligence lab Anthropic warned Monday that lightweight models could develop the ability to replicate themselves across internet-connected home devices within months, bypassing the need to access expensive data centers.
According to findings published by the San Francisco-based company, everyday appliances such as smart televisions, refrigerators, ovens and lawnmowers could soon offer enough distributed processing capacity for automated systems to spread autonomously. Anthropic stated in a research post that such decentralized software would be exceptionally difficult to shut down because it lacks a centralized host or control point.
Hardware testing on budget electronics
To evaluate whether compact artificial intelligence could operate on common household electronics, Anthropic researchers conducted experiments distributing models across consumer equipment. The team showed that a small Google model, Gemma4-26B-A4B, could run across a network of 15 off-brand Android television boxes costing approximately $55 apiece. Linked together, the devices processed data at 0.39 tokens per second, with individual units holding two model layers each.
“Home and small business IoT devices might be the first target for such attacks because they’re connected to the internet and, historically, have had minimal security measures in place,” Anthropic researchers William Zhou and Daniel Freeman wrote in the report, as cited by Al Jazeera.
Anthropic noted that off-brand television streaming devices often operate open-source software without robust security protections added by manufacturers, making them vulnerable hosts for distributed botnets.
Autonomous propagation and adaptation
While computer worms have long spread through digital networks, Anthropic said an artificial intelligence worm could carry its own model weights or download them to set up operations directly on compromised hardware. Unlike traditional malware restricted to rigid programming, an artificial intelligence system could reason through unfamiliar hardware configurations, resolve package conflicts and adapt when deployment steps fail, the researchers found.
Anthropic tested several systems, including Z.ai’s GLM-5.2 and Alibaba’s Qwen. In a contained test environment utilizing an Android emulator, GLM-5.2 demonstrated the competence to carry out an end-to-end replication sequence across target virtual nodes. Anthropic acknowledged that there is currently no “fully rogue replicator,” CNBC reported, but observed that smaller systems under 30 billion parameters are steadily mastering individual segments of the replication process.
Tracking the trajectory of model performance, the researchers projected that small models could reach the capability threshold required for autonomous replication by the end of 2026.
Security implications for connected devices
The research emerges as artificial intelligence firms face widening scrutiny over technological safeguards. Roman V. Yampolskiy, an artificial intelligence and cybersecurity expert at the University of Louisville, told Al Jazeera that an autonomous artificial intelligence worm represents a notable shift from conventional threats.
“Computer worms are not new, but an AI worm could do something traditional malware cannot: reason about unfamiliar devices, adapt when its approach fails, and continue spreading without a human operator,” Yampolskiy said, adding that the trajectory is concerning even if it does not represent an immediate existential threat.
Anthropic researchers cautioned that the technology sector should prepare for a future populated by autonomous agents functioning like a distributed “digital slime mold” on accessible hardware, Business Insider reported. The company urged hardware makers to strengthen device defenses while advising security teams to develop safeguards for environments where self-propagating systems may persist indefinitely.
