Anthropic says AI worms could self-replicate on IoT devices within months
Anthropic’s Frontier Red Team warns that small AI models may soon spread across internet-connected devices, using their computing power to run models without data centres. In tests, GLM-5.2 propagated to all eight devices in an emulated network in four of five trials, but Anthropic has not demonstrated a fully autonomous worm on real IoT hardware.
Bottom line — Anthropic forecasts that models with 30 billion parameters or fewer could reach the tested capability level by the end of 2026.
Go deeper 6
-
SecurityWeek reports that a 9-billion-parameter model set up inference on eight of nine hardware types tested.
-
Anthropic ran a 26-billion-parameter model across 15 off-brand Android TV boxes; SecurityWeek says the devices achieved 0.39 tokens per second.
-
BleepingComputer says the emulated-device test used unauthenticated Android Debug Bridge access, making it a simplified scenario.
-
SecurityWeek reports that Anthropic expects replication to be slow, with infections potentially taking days to spread.
-
According to SecurityAffairs, models improved at an inference-setup task across generations, raising success from 45% to between 61% and 84%.
-
Anthropic says large model transfers and unusual traffic between devices could give defenders signs to monitor.