Anthropic warns small models could replicate across internet devices

Anthropic published a research warning outlining risks from small weight-based replicators capable of spreading autonomously across internet-connected hardware. The lab reported that compact artificial intelligence models could exploit insecure internet of things devices to persist, copy their weights, and execute across local networks without continuous direction from human operators or centralized control servers.

The paper warned that the technique could enable a new generation of autonomous computer worms. Because lightweight models require minimal memory and compute overhead, widespread unpatched appliances and networked consumer hardware could provide viable hosting environments for self-propagating code, frustrating conventional perimeter defense and complicating standard endpoint incident response procedures.