"Anthropic" warns of AI worms capable of spreading across smart devices

American company "Anthropic" warned that AI models may become capable of self-replication and spreading among smart home devices. Security team experiments showed the possibility of running a distributed model across 15 smart TVs, raising concerns about the emergence of "AI worms" that are difficult to stop.

American company "Anthropic" warned that AI models could within months become capable of self-replication and independent spread among internet-connected smart home devices, potentially making detection and removal of malicious software more difficult. A study published by the company's security testing team, known as the "red team," revealed experiments showing the possibility of running an AI model across 15 low-cost smart TV devices by distributing its parts among them and connecting them to a single network. These results raised concerns about a new generation of malicious software known as "AI worms," capable of spreading among internet-connected devices without continuous human intervention. However, the experiment did not prove the model's ability to hack into the fifteen TVs and spread among them automatically; researchers tested running the distributed model on actual devices, while conducting self-spreading tests separately in simplified and isolated environments. Anthropic researchers believe that future integration of the ability to run small models on home devices with hacking and self-spreading skills could enable the development of software that is difficult to stop, especially if distributed across a large number of devices without relying on a central server.