Anthropic researchers warn that improving small AI models could soon form self-replicating worms across insecure household devices, though current systems remain unreliable.