Anthropic warns AI agents could spread through connected appliances

Anthropic engineers warn that small but capable artificial intelligence models could soon keep themselves running and replicate on internet-connected devices such as robot vacuums, washing machines, refrigerators and televisions. In a technical commentary by Kristjan Port, ERR’s R2 Portal reports that the milestone could be only months away in the worst case.

Anthropic’s warning adds a new dimension to the risks posed by AI agents, the commentary says. It points to the fear of an intelligent threat that is hidden and beyond control.

Water can find every crack that a builder failed to protect, but that does not make it intelligent, Port writes. A different kind of fear is evoked by the unseen intelligent creature moving through ventilation ducts and behind walls in Ridley Scott’s film “Alien”. Though fictional, the predator left many viewers terrified, with some screaming or rushing out of cinemas. The film was released in the late 1970s, during a period of anxiety about secret government plans to control the public and corporate schemes driven by greed and power.

Books can also have a powerful effect. In the mid-1990s, Richard Preston’s nonfiction book “The Hot Zone” described dangerous viruses, including Ebola and Marburg, and the damage they can cause in the body. Horror writer Stephen King called its opening chapter “one of the scariest things I’ve ever read.”

Public fear prompted government responses. The US Centers for Disease Control and Prevention changed its activities, while the World Health Organization ran campaigns to raise awareness of viral threats and funded prevention projects. The book was later criticised for its dramatised presentation.

Against that backdrop, Anthropic says some of its newer models could soon become “self-replicating” and find ways to enter devices connected to the internet.