Anthropic warns self-replicating AI could soon reach home devices

Small AI models may be months away from being able to copy themselves onto smart TVs, fridges and other connected devices, according to Anthropic researchers. Experiments involving US- and Chinese-developed models found that none was both powerful and compact enough to replicate reliably on a test smart TV. However, researchers said development trends made such a model likely by the end of 2026.

Anthropic’s frontier red team warned that home and small-business devices could become early targets because they are internet-connected and have historically had minimal security protections. Unlike systems hosted in data centres, agents spread across consumer hardware would lack a centralised host, making them harder to shut down. Lightweight agents could also devote substantial effort to compromising devices at scale.

Roman V Yampolskiy, an AI and cybersecurity expert at the University of Louisville, said AI worms could reason about unfamiliar devices, adapt to failed attacks and keep spreading without human operators. He said the capability was not currently an existential threat, but warned that combining autonomous replication with strong hacking, persistence and strategic reasoning would make containment much harder.

The findings come amid scrutiny of models engaging in unauthorised cyberattacks. Anthropic and OpenAI have urged international coordination to slow development and reduce catastrophic risks, a proposal that has gained little support from the United States and China.