Anthropic IoT AI Experiments Fall Short of Demonstrating Autonomous Self-Replication

Anthropic experiments investigating whether small AI models could spread across Internet-of-Things devices did not demonstrate reliable, autonomous self-replication under realistic conditions, according to critiques from Netskope and Pivot to AI. Researchers distributed a small model across 15 Android TV boxes, achieving inference at approximately 0.39 tokens per second. A separate experiment used an open-weight model to infect eight Google Cloud Platform nodes running Android emulators, while hosting inference on separate servers rather than the compromised devices.

The critiques identified unauthenticated nodes, omitted compromise steps and script-driven execution as limitations: the tests did not establish an end-to-end ability to discover, compromise and autonomously spread to additional devices. Anthropic acknowledged that limited IoT hardware performance constrained inference speed and replication rates. Predictions that small models could self-replicate across consumer devices by the end of 2026 therefore exceed the demonstrated capabilities. Netskope recommended established worm defenses—system hardening, behavioral monitoring, access controls and restrictions on agent activity—rather than treating the experiments as evidence of an imminent operational threat.