Anthropic researchers William Zhou and Daniel Freeman report tests of AI replication without a datacenter on October 12. One model ran across 15 purchased Android TV boxes.
In a simplified network trial, GLM-5.2 propagated a Llama3.1-8B payload and set up inference. The eight firewalled emulator nodes offered unauthenticated access; the controlling model ran on a separate server.
The authors expect smaller models could reach these capabilities. That is a forecast, not an observed escape: the tests simplify access and do not demonstrate a complete replicator running independently on third-party devices.
They suggest network monitoring, quarantine and safer device defaults as possible defenses.