ANTHROPIC'S OWN RED TEAM RAN AN AI ON 15 $55 TV BOXES — WARNS AI WORMS ARE 'MONTHS AWAY'

Anthropic's Frontier Red Team published research on autonomous replication and adaptation, testing whether small models can copy themselves across poorly secured consumer devices without human help. Such an AI worm would carry its weights, or re-download them from a public hub, and run inference on the hardware it compromises. To test whether cheap hardware is enough, researchers bought 15 Android TV boxes costing about $55 each, with 4GB of RAM, and ran the Gemma4-26B-A4B model distributed across them, at about 0.39 tokens per second.

Most models could complete individual steps such as scanning, gaining access and setting up an inference environment, but struggled to chain them together. Only the open-weight GLM-5.2 completed the full chain in the simplified test. Anthropic projects that if small models keep improving at the current rate, many could reach that level by the end of 2026. A spread-out worm would have no central server to shut off, though it leaves a visible network signature and quarantine cleared cited infections within 14 days.