Japanese note explains Anthropic's self-replication evaluation

What the research found

Japanese note.com article summarizes Anthropic's October 12, 2026 evaluation of whether small models can replicate across IoT and consumer devices; it is a threat-model test in isolated environments, not evidence of AI worms in the wild.

Claims and how well the evidence supports them

1. Anthropic published the self-replication evaluation on October 12, 2026 (per article; jev did not confirm the date) 92%
2. A 9B-scale model ran inference on 8 of 9 tested hardware types 98%
3. Gemma 4 26B-A4B split across 15 Android TV boxes with 4GB each ran at 0.39 tokens per second 98%
4. This is an isolated evaluation, not confirmation of AI worms on the real internet 99%

Evidence

• note.com https://note.com/turuto/n/n4f9f5443cc62?hl=en

Actionability rated medium.