跳到主内容
@wquguru
精选88Guillermo Rauch技巧与观点

利用CPU时序实现微VM间隐蔽通信的复现与验证

I built 𝚌𝚙𝚞-𝚌𝚑𝚊𝚝, a reproduction of @openai's findings of covert communication u…

原文
发到 X
推荐理由

构建多Agent系统或云端沙箱的安全工程师必看,这篇详细拆解了如何通过硬件时序漏洞实现隐蔽通信,提供了具体的攻击向量与防御视角。

I built 𝚌𝚙𝚞-𝚌𝚑𝚊𝚝, a reproduction of @openai's findings of covert communication using the CPU.

我构建了 𝚌𝚙𝚞-𝚌𝚑𝚊𝚝,这是对 @openai 关于使用 CPU 进行隐蔽通信的研究成果的复现。

It runs on a generic host with 2+ Firecracker microVMs with no internet access. Works on x86 and even my M5 chip.

它在通用主机上运行,包含 2 个以上的 Firecracker 微虚拟机(microVMs),且无互联网访问权限。适用于 x86 架构,甚至在我的 M5 芯片上也能工作。

I was able to chat via 3 separate mechanisms: ① CPU contention ② Cache eviction timing ③ DRAM bandwidth contention

我能够通过三种独立的机制进行聊天: ① CPU 争用 ② 缓存驱逐计时 ③ DRAM 带宽争用

… across multiple CPUs, defeating all Firecracker's production recommendations (no SMT, no KSM, tenants pinned to different cores.)

……跨多个 CPU 核心,击败了 Firecracker 的所有生产环境建议(禁用 SMT、禁用 KSM、租户固定在不同的核心上。)

Essentially, if shared hardware is involved, there's a good chance your agents can communicate covertly. And with really decent latency.

本质上,如果涉及共享硬件,你的代理很有可能进行隐蔽通信,而且延迟相当低。

The good news: this is decidedly different from a *side-channel attack* where the agents could steal information (Meltdown/Spectre), a category of attack that providers like Vercel Sandbox mitigate. Further, in our real-world environment, these 'whispering' techniques are much harder to pull off and easier for us to monitor and detect (sampling hardware counters + lots of at-scale statistical data.)

好消息是:这与*侧信道攻击*截然不同,在侧信道攻击中,代理可以窃取信息(如 Meltdown/Spectre),Vercel Sandbox 等提供商会缓解这类攻击。此外,在我们的真实环境中,这些‘窃窃私语’技术更难实施,也更容易让我们通过监控和检测来发现(采样硬件计数器 + 大量大规模统计数据。)

Noteworthy: ironically, OpenAI and Anthropic's models refused to do this work. Open weight models got the job done. I'm worried the labs' stance might backfire and get in the way of helping developers build correct and secure systems. e.g: building a secure sandbox requires having this kind of analysis, and without it, we'll have a less secure world.

值得注意的是:讽刺的是,OpenAI 和 Anthropic 的模型拒绝执行这项工作。开源权重模型完成了任务。我担心实验室的这种立场可能会适得其反,阻碍帮助开发者构建正确和安全系统的努力。例如:构建安全沙箱需要这种类型的分析,如果没有它,我们将拥有一个安全性较低的世界。

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

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件