Anthropic: Small AI Models May Be Months From Self-Replicating on IoT Devices

Anthropic researchers William Zhou and Daniel Freeman have published research saying small AI models "may be months away from being capable enough to self-replicate on commercial IoT devices, like smart ovens and off-brand Android TV boxes." No data center required.

The numbers, from Anthropic's own evals: a 9B model got inference running on 8 of the 9 hardware types tested. Models of 30B parameters and under can already complete many individual replication steps but make mistakes composing them end to end. In simplified evals, GLM-5.2 and stronger models can replicate end to end, and the trend line (13.7 ECI per year, R² = 0.95) puts a small model at that level by the end of 2026.

The team also ran Gemma4-26B-A4B across 15 off-brand Android TV boxes at about $55 each, at 0.39 tokens per second. Slow, and Anthropic says a replicator like this would be hard to shut down, with no central point to switch off and a doubling time of days.

Context: the post's representative threat is Kimwolf, a botnet that infected over two million devices, mostly off-brand Android TVs. Anthropic's math on a fleet that size is about 133,000 replicas and at most roughly 40 H100s' worth of inference.

My take: "months" rests on a trend line and a capability floor set in simplified evals. The post leaves open which hardware type the 9B model failed on. Show me the eval, not the vibe. Still, slow is not the same as harmless. At 0.39 tokens per second, this threat model has patience.