Anthropic researchers warn that small rogue AI models may be months from autonomously replicating across poorly secured internet-connected devices, including Android TVs, cameras and smart appliances. Such networks could spread without human intervention and have no central shutdown point.