Anthropic has raised concerns that small AI models may be capable of self-replicating across IoT devices by the end of this year, spreading in a manner similar to computer worms—malware that propagates automatically without human intervention. This threat becomes more challenging to contain because the models would execute locally on devices rather than running from centralized cloud servers.
In proof-of-concept tests, researchers successfully distributed the Gemma 4-26B model across 15 Android TV devices costing approximately $55 each, each equipped with 4 GB of RAM. However, the model responsible for orchestrating the propagation remained on a separate server. Anthropic emphasizes that researchers have not yet successfully simulated a fully autonomous, uncontrolled replicator in their experiments, suggesting the technology remains in early stages but warrants careful monitoring.