Anthropic warned in October 2026 that an AI model self-replicating across IoT devices would be hard to shut down because it would have no central point to switch off