Anthropic tests AI model across 15 Android TV boxes

Anthropic said research released on Monday 12 October found that AI models may be nearing the technical ability to establish themselves on different computer hardware, reproduce their operations and spread through networks. The company said the work did not demonstrate an uncontrolled AI system spreading independently across real-world networks.

Researchers William Zhou and Daniel Freeman tested an AI model across 15 Android TV boxes costing about US$55 each, Anthropic said. The model’s computational work was distributed among the devices, showing that the equipment could collectively run the system under controlled conditions with simplified security barriers.

Anthropic said smart televisions, Android TV boxes, security cameras and other internet-connected appliances could be environments where autonomous AI programs operate. It said such systems could adapt to unfamiliar devices, unlike traditional malicious software written for specific systems.

The company said internet service providers could identify unusual network activity because a self-replicating AI system would need to move substantial data and communicate between devices. It said manufacturers could reduce vulnerabilities through stronger security, removal of insecure default passwords and software updates.

Anthropic estimated that smaller AI models could reach the capability demonstrated by the larger model before the end of 2026 if their improvement continues at its historical pace.