Anthropic’s Frontier Red Team published research warning that small open-weight AI models could, within months, be able to copy themselves onto poorly secured connected gadgets like smart ovens and cheap Android TV boxes. Such a replicator would likely spread slowly, but Anthropic said it would be hard to stop because there’s no central point to shut off, and unlike conventional malware it could work out how to run on unfamiliar hardware by itself. Models of 30 billion parameters or fewer can already handle many of the steps but still make mistakes chaining them together. The only open-weight model to replicate end to end was the larger GLM-5.2, in a simplified test where the devices’ Android debugging port required no login, which Anthropic called an “upper bound on real-world success.”
