谷歌于10月22日描述了Agent Compute Engine架构,该方案为每个AI会话提供独立的microVM以实现硬件级隔离,且不具备网络接口卡 1。针对这一描述,一位开发者使用约2200行Go代码和Firecracker v1.17.0构建了原型系统,以验证该架构的可行性与性能表现 1。在裸金属环境(Intel i7-6700K)下的测试显示,利用快照分叉技术,64个空闲虚拟机每个仅消耗主机约3 MiB内存,且快照恢复延迟仅为25ms,比冷启动快26倍 1。然而,在Mac mini (Apple M4) 的嵌套虚拟化环境中,12个同时启动的分叉均无法在30秒内就绪,暴露出该技术路线在扩展性上的缺陷 1。
在安全模型方面,该实现让Guest环境完全无网卡,流量通过vsock转发至主机代理,真实凭证仅在主机网关处注入,从而确保Guest中不包含敏感Token 1。文章对比了该技术与容器及gVisor的安全性差异,并指出Alibaba OpenSandbox和Meta/HuggingFace OpenEnv等也是该领域的执行环境项目 1。此外,开发者强调了潜在风险,包括共享快照页面可能构成的侧信道攻击面,以及用户态随机数需在Fork后重新播种以避免重复序列的问题 1。
A developer has recreated and tested the AI agent microVM isolation architecture recently described by Google, utilizing approximately 2,200 lines of Go code built on Firecracker v1.17.0 1. This implementation aims to replicate the security model of Google's Agent Compute Engine, which was outlined on October 22 and features a dedicated microVM for each AI session with hardware isolation and no network interface card 1. The prototype employs a security framework where the guest environment lacks a NIC, routing traffic through vsock to a host proxy, while real credentials are injected only at the host gateway to ensure sensitive tokens remain outside the guest environment 1.
Performance benchmarks reveal significant efficiency gains on bare-metal hardware but highlight scalability issues in nested virtualization environments 1. On an Intel i7-6700K bare-metal setup, each of the 64 idle forked VMs consumed only about 3 MiB of host memory, and snapshot restoration latency was measured at 25ms, making it 26 times faster than cold starts 1. In contrast, testing on a Mac mini equipped with an Apple M4 chip under nested virtualization exposed critical bottlenecks, as none of the 12 simultaneously started forks were ready within a 30-second window 1. The article further notes potential security risks, such as shared snapshot pages creating side-channel attack surfaces and the need to reseed user-space random number generators after forking to prevent duplicate sequences 1. Additionally, the project draws comparisons with other execution environment initiatives like Alibaba OpenSandbox and Meta/HuggingFace OpenEnv 1.
评论
还没有评论,欢迎留下第一条。