Anthropic's Frontier Red Team published experiments suggesting that small AI models may be months away from being able to copy themselves onto cheap IoT devices. In tests, a 9-billion-parameter model got inference running on 8 of 9 hardware types, while a larger model ran distributed across 15 Android TV boxes costing about $55 each. Stronger models completed full replication cycles in a simplified setting. The company stresses that no working rogue replicator exists yet.
What it means: The research does not describe a threat that exists today, but it shows where risk is heading: away from expensive data centres and towards cheap, poorly protected devices found in every office and shop. Cameras, smart TVs, thermostats and point-of-sale gadgets often run old software with factory passwords. If a future program carrying its own model can work out how to use unfamiliar devices, old-style botnets gain new flexibility. For Greek SMEs the practical answer is well known but rarely applied: a separate network for IoT devices, changed passwords, remote access switched off where it is not needed and hardware bought from vendors who ship updates. The EU Cyber Resilience Act will gradually push such requirements onto manufacturers, yet equipment already installed will stay in use for years. Companies building devices or firmware, such as ordering kiosks, should design on the assumption that someone will try to run foreign code on them.
Read the original on Anthropic
Related service: Custom software