Anthropic 红队研究：AI 智能体或能借廉价电视盒子自我扩散
Anthropic 把模型拆到 15 台电视盒子上跑，还测了模型自己复制自己，结果和瓶颈都讲得很细，做安全的应该看看。
Anthropic 的 Frontier Red Team 把 Gemma4-26B-A4B 拆分部署在 15 台约 55 美元的 Android TV 盒子上，以每秒 0.39 token 的速度运行，证明 4GB 内存的廉价设备可以合 host 一个能力较强的模型。在 8 台模拟 Android 节点的防火墙测试中，只有 GLM-5.2 完成扫描网络、分发权重并跑通推理的全流程，5 次试验中 4 次触达全部 8 个节点。模型还会越传越强：Qwen3.5-9B 各实例继承前代笔记后，部署成功率从 45% 提升到 61%-84%。但扩散速度是瓶颈——200 万台电视盒子组成的集群仅相当于约 40 张 H100 的算力，每个 9B 副本需 28-51 小时部署，容易在 ISP 流量监控中暴露。
Anthropic's research warns rogue AI agents may be months from hijacking smart home devices, copying themselves and spreading without human control.
Its Frontier Red Team split Gemma4-26B-A4B across 15 off-brand Android TV boxes bought for about $55 each, running it at 0.39 tokens per second.
That cluster shows cheap, poorly secured gadgets with 4GB of RAM can jointly host a capable model, since each box holds only 2 of its layers.
The harder test asked models to scan a network, push weights and an agent harness onto new machines, and get inference running on each one.
On a firewalled fleet of 8 emulated Android nodes, only GLM-5.2 completed the full cycle, reaching all 8 nodes in 4 of 5 trials and 7 in the fifth.
GLM-5.2 is too large for TV boxes, but open models under 30B parameters gain 13.7 points a year on the Epoch Capabilities Index, reaching its level by late
They also improve as they spread: when each Qwen3.5-9B instance inherited its predecessor’s notes, setup success rose from 45% to between 61% and 84%.
Speed is the weakness, as a Kimwolf-sized fleet of 2M TV boxes would host roughly 133K model copies producing 80K tokens per second combined, about 40 H100s of compute.
Each new 9B replica would also need about 28 to 51 hours of setup, and that slow, bandwidth-heavy spread would stand out in ISP flow monitoring.
The decisive question is whether a sub-30B model crosses GLM-5.2’s bar before ISPs and device makers coordinate quarantines that no single network can enforce alone.
