Anthropic researchers say small AI models could copy themselves across poorly secured smart TVs, fridges and other connected devices without data centres, making them difficult to shut down. Tests found no current model capable of doing so reliably, but one could emerge by the end of 2026.