Anthropic warns small AI models may soon self-replicate across insecure devices

Anthropic’s findings, tested in emulated networks, show models hopping between devices and improving over time via sharding. Experts question whether this qualifies as true self-replication, comparing it to malware droppers that install coordination tools rather than copying code. Both Canal and Bailey emphasize that existing security practices—patching devices, enforcing isolation, and strong authentication—remain the best defenses, despite AI’s potential to lower the barrier for attackers. Anthropic also highlights regulatory measures like the UK’s PSTI Act, which bans weak IoT passwords, as a model for mitigation.