Anthropic warns AI models could self-replicate across home devices within months

Anthropic has warned that small AI models could soon copy themselves across internet-connected home devices, in a report published Monday.

• Anthropic said small AI models "may be months away from being capable enough" to self-replicate on internet-of-things devices such as smart TVs and fridges.
• Tests by the company's frontier red team, including running Google's Gemma4-26B-A4B model across 15 Android televisions, determined such systems would be challenging to shut down without a centralised host.
• Researchers predicted a reliably self-replicating model could emerge by the end of 2026, targeting home and small-business devices with minimal security.
• Anthropic separately announced measures that reduced illicit extraction of its models' data by 90–99 percent.
• In its report, Anthropic predicted that future attacks would involve swarms of self-replicating AI models.
• The report listed ovens among the targeted smart-home devices, in addition to TVs and refrigerators.