Anthropic: small AI models can self-replicate across smart devices
Anthropic's Frontier Red Team said on October 12 that models under 30B parameters already handle individual self-replication steps, while GLM-5.2 completed the full cycle once safety barriers were removed. In one test a 26B model was spread across 15 $55 Android TV boxes; no real-world home cases have been reported yet.
- Models under 30B parameters handle self-replication steps individually
- GLM-5.2 completed the full replication cycle with safety barriers removed
- A 26B model was spread across 15 $55 Android TV boxes
- Anthropic expects full capability before the end of 2026
Read next
Security