Anthropic warned that compact artificial-intelligence models may soon gain the ability to self-replicate across internet-connected devices, forming autonomous botnets that are difficult to disrupt. The San Francisco–based AI lab said so-called “AI worms” could copy their model weights onto compromised hardware—such as low-cost Android TV boxes—or re-download them from public hubs, then run inference locally and propagate further, eliminating any centralized kill switch.
In tests of 11 open-weight models, Anthropic found one configuration in which GLM-5.2 reliably moved Llama 3.1-8B across eight Android emulator nodes in most trials. The company outlined multiple replication paths, including weight-based spread on insecure IoT gear, stealth deployments in high-compute environments, and abuse of stolen API credentials—though anomalous API usage may be easier to detect.
While Anthropic says a fully autonomous “rogue replicator” has not yet emerged, the building blocks are falling into place. The prospect raises security and compliance concerns for device makers, cloud providers and enterprises, which may need to harden IoT endpoints, monitor unusual compute patterns and tighten API controls as attackers experiment with distributed AI inference at scale.
Related article:
