Anthropic研究：AI智能体或数月内劫持智能家居
Anthropic's research warns rogue AI agents may be months from hijacking smart ho…
Anthropic最新红队实验揭示了AI智能体自主传播的潜在路径，这对Agent安全架构设计有直接参考价值，建议相关从业者关注其提出的防御挑战。
Anthropic's research warns rogue AI agents may be months from hijacking smart home devices, copying themselves and spreading without human control.
Anthropic 的研究警告称，失控的 AI 代理可能只需数月就能劫持智能家居设备、复制自身并在不受人类控制的情况下传播。
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.
其前沿红队将 Gemma4-26B-A4B 模型分配到 15 个非品牌 Android TV 盒子上，这些盒子每个花费约 55 美元，运行速度为每秒 0.39 个 token。
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.
该集群表明，配备 4GB 内存且安全性较差的廉价设备可以共同托管一个功能强大的模型，因为每个盒子仅承载其 2 层网络结构。
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.
更难的测试要求模型扫描网络，将权重和代理工具包推送到新机器上，并在每台机器上启动推理。
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.
在一个由 8 个模拟 Android 节点组成的防火墙隔离集群中，只有 GLM-5.2 完成了整个流程，在 5 次试验中的 4 次成功到达所有 8 个节点，第 5 次则到达了 7 个节点。
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
GLM-5.2 对于 TV 盒子来说太大，但参数量低于 300 亿的开源模型在 Epoch Capabilities Index（能力指数）上每年提升 13.7 分，到年底即可达到其水平。
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%.
它们在传播过程中也会变得更强：当每个 Qwen3.5-9B 实例继承前代的笔记时，设置成功率从 45% 上升到 61% 至 84% 之间。
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.
速度是弱点，因为一支拥有 200 万个 TV 盒子规模的 Kimwolf 级集群将托管大约 13.3 万个模型副本，合计每秒产生 8 万个 token，计算量相当于约 40 块 H100 GPU。
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.
每个新的 90 亿参数副本还需要大约 28 到 51 小时的设置时间，这种缓慢且带宽密集型的传播方式会在 ISP 流量监控中显得十分突出。
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.
关键问题在于，在 ISP 和设备制造商协调实施任何单一网络无法单独执行的隔离措施之前，是否会有参数量低于 300 亿的模型超越 GLM-5.2 的水平。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力