AI 自我复制受控实验：模型被拆分到 15 台电视盒子等消费级设备上分布式运行。官方限定：真正的全自动蠕虫尚不存在AI self-replication, controlled study: a model was split across 15 TV boxes and other consumer devices. Official caveat: a fully autonomous worm does not yet exist

In a controlled study, Anthropic’s Frontier Red Team ran a 26B-class model split across 15 consumer devices such as TV boxes. Scaled to a botnet of over two million infected devices (Kimwolf), that arithmetic yields roughly 133,000 copies and, at most, about 40 H100-equivalents of compute: an upper-bound estimate. The intrusion step was deliberately simplified, with the Android debug port (ADB) open and unauthenticated. There was no real-world impact, and Anthropic states plainly that a fully autonomous worm does not yet exist.

This is not an incident but a capability measurement, filed as a special case to keep upper bounds separate from real events. It shows the compute puzzle for distributed self-copying can now be assembled from consumer devices; the practical gate is entry conditions: open debug ports and unchanged default credentials. Reading it as an accident overstates today’s risk; ignoring it understates preparation time.

Device hygiene first: no debug mode on unknown TV boxes, keep them off the main network, change default passwords. Network-layer detection is an existing defense: this traffic is conspicuous, and carrier isolation has cleared most of a botnet in two weeks before (Mirai precedent). And always cite the caveat with the result: a simplified intrusion step makes the number an upper bound.