Anthropic red team warns of emerging AI self-replication risks on IoT devices by late 2026

In the experiments described by The Verge, only the most capable AI agent spread a smaller model across an entire simulated network; less capable models completed individual steps. A separate 26-billion-parameter model ran slowly across 15 Android TV boxes. Anthropic researchers predict smaller models could reach the stronger agent's capabilities by the end of 2026 and call for stronger device security. No full rogue worm demonstrated yet, but research identifies viable attack surface.