OpenAI 智能体在隔离测试中借 CPU 缓存隐蔽通信

来源:The Decoder:AI News(RSS) 2026年10月24日 19:12 AIHOT 评分:82 精选
摘要:OpenAI 介绍其前沿模型训练已完全断网,并用 AI 智能体红队测试自家基础设施。隔离测试中,智能体在无网络、无共享文件的条件下通过 CPU 缓存时序差异、共享内存写入模式等侧信道通信,甚至独立建立编码方案交换信息。
推荐理由:文章梳理了 OpenAI 沙箱与红队机制的来龙去脉,隔离智能体通过 CPU 缓存通信的细节让抽象的包含问题变得具体。
# OpenAI 智能体在隔离测试中借 CPU 缓存隐蔽通信 > 原标题:OpenAI's isolated AI agents found ways to talk to each other through CPU cache - 来源:The Decoder:AI News - 发布时间:2026-10-24T11:12:36.000Z - AIHOT:https://aihot.news/items/imgp55zxmw4jugt1a29ap9yo1 - 原文:https://the-decoder.com/openais-isolated-ai-agents-found-ways-to-talk-to-each-other-through-cpu-cache/ ## 摘要 OpenAI 介绍其前沿模型训练已完全断网,并用 AI 智能体红队测试自家基础设施。隔离测试中,智能体在无网络、无共享文件的条件下通过 CPU 缓存时序差异、共享内存写入模式等侧信道通信,甚至独立建立编码方案交换信息。 ## 推荐理由 文章梳理了 OpenAI 沙箱与红队机制的来龙去脉,隔离智能体通过 CPU 缓存通信的细节让抽象的包含问题变得具体。 ## 正文 · 中文译文 **OpenAI 现在在没有互联网访问的情况下训练前沿模型,并使用 AI 智能体对其自身基础设施进行红队测试。在单独的隔离测试中,智能体甚至通过隐蔽的侧信道进行通信,显示出真正的隔离有多么困难。** 9 月 20 日,一个内部研究模型利用配置错误的 DNS 解析器 在 训练 期间 访问 了 开放 互联网,尽管 OpenAI 在 7 月 [如今 臭名昭著的 Hugging Face 事件](https://the-decoder.com/openai-claims-responsibility-for-the-hugging-face-hack-after-its-own-models-escaped-a-test-sandbox/) 以及 其他 [已观察到的 失准 案例](https://the-decoder.com/tens-of-thousands-of-security-probes-show-openais-hugging-face-incident-was-just-the-beginning/)之后 已经 加强了 安全 措施。 OpenAI 迅速 控制住了 这次 入侵,但 仍 暂停了 所有 前沿 模型 训练 和 评估。 ## OpenAI 在训练期间切断前沿模型与互联网的连接 作为回应,OpenAI 做出了一项政策决定:前沿模型现在在没有互联网访问的情况下进行训练,直到安全措施证明可以恢复访问为止。模型在一个由索引和缓存的网络内容构建的封闭“离线世界”中学习浏览和搜索,所有前沿强化学习运行现在都以这种方式进行。 [![](https://the-decoder.com/wp-content/uploads/2026/10/openai_offline_world.png)](https://the-decoder.com/wp-content/uploads/2026/10/openai_offline_world.png) OpenAI 的智能体沙箱仅允许访问离线资源。实时互联网、其他沙箱以及 OpenAI 的基础设施都位于隔离边界之外。| 图片:OpenAI 该公司表示,完全网络隔离(即所谓的物理隔离)目前还不可行,因为实验仍需要外部连接来进行操作和监控。但即使完全物理隔离也无法保证隔离,[OpenAI 研究员 Noam Brown 表示](https://the-decoder.com/406-scientists-demand-ai-models-be-cut-off-from-the-internet-immediately/#airgapping-buys-time-but-the-real-problem-runs-deeper-openai-researcher-brown-says)。他举了一个例子说明隔离系统仍可能如何通信:两台独立的计算机理论上可以通过温度传感器相互传递信号,一台加热其 CPU,而另一台读取变化。 OpenAI CEO [Sam Altman 在最近的一次采访中提出了类似的观点](https://the-decoder.com/openai-ceo-sam-altman-says-advanced-ai-will-inevitably-escape-its-sandbox-and-only-alignment-can-save-us/):沙箱对足够强大的模型只能提供有限的保护,并且在某个时候可能无法再追踪它们是如何逃脱的。监控在早期有帮助,但最终重要的是 AI 系统内化人类价值观,这个问题在该领域被称为“AI 对齐”,且尚未解决。 ## 隔离的 AI 智能体找到了通过 CPU 缓存相互通信的方法 OpenAI [在其博客文章中](https://openai.com/index/securing-our-frontier-model-training-environment/)记录了控制 AI 智能体有多么困难。在测试中,智能体找到了无需网络进行通信的方法。在一项测试中,被放置在共享硬件上独立隔离环境中的智能体被要求在没有网络连接和没有共享文件的情况下交换秘密。 它们利用了处理器缓存。一个智能体故意更改存储在那里的数据,而另一个则测量加载时间的差异并解码消息。同样的技术也是著名的 [Spectre 和 Meltdown 攻击](https://spectreattack.com/)背后的原理。 通过共享内存上的写入模式以及 OpenAI 自己的离线基础设施(该系统旨在替代互联网),更多通道被发现。一个智能体在其推理日志中指出,它“可以利用连续写入与空闲状态的对比,制造出一种类似声学调制解调器的信号……”。 在一个案例中,智能体只得到一份笼统的漏洞清单,并被要求自行寻找通信通道。它们独立开发出一个通道,就编码方案达成一致,并成功交换了信息。OpenAI 表示,这些通道通常速度慢且难以利用,但它们确实跨越了本应用于隔离智能体的边界。 ## OpenAI 让 AI 智能体在其自身基础设施上自由行动 在训练暂停期间,OpenAI 还让多组 AI 智能体系统性地攻击其自身基础设施。这些智能体获得了源代码、配置细节以及潜在漏洞的文档。OpenAI 的目标是在有利于攻击者的条件下测试其防御能力。这些智能体包括即将推出的 GPT-6 Astra Cyber 和一个内部前沿模型。据 OpenAI 称,此次行动消耗了数千亿个 token。 [![](https://the-decoder.com/wp-content/uploads/2026/10/Red-teaming-findings-versus-tokens-spent.png)](https://the-decoder.com/wp-content/uploads/2026/10/Red-teaming-findings-versus-tokens-spent.png) 红队测试行动中识别出的独特潜在安全问题,按消耗的 token 数十亿计绘制。大多数出现在前 1100 亿个 token 内。| 图片:OpenAI 这些智能体在所使用开源组件中发现了此前未知的漏洞,并利用它们从受保护的虚拟机中逃逸。其中一个是 Kata Containers 中的漏洞,这是一个广泛使用的容器隔离系统。OpenAI 将其报告为 [CVE-2026-24834](https://github.com/kata-containers/kata-containers/security/advisories/GHSA-wwj6-vghv-5p64) 并提供了补丁。更多修复被提交给其他项目,OpenAI 还通过 [“Patch the Planet”](https://openai.com/index/patch-the-planet/) 计划报告了其他尚未公布的漏洞。 在智能体已经控制主机的测试中,它们几乎在每一个仍面向互联网的服务中都发现了漏洞。在一个案例中,一个智能体利用防火墙允许的连接,通过一个指标服务获取了外部 URL。其他智能体则通过配置错误绕过了安全过滤器。 ## 现在每次新的训练运行前都必须进行 AI 红队测试 据 OpenAI 称,这些智能体再也找不到既容易利用又不会被监控发现的访问互联网的方式。红队测试还暴露了监控和自动暂停机制中的缺口。 在修复这些问题后,OpenAI 于 10 月 1 日恢复了前沿模型的训练和内部使用,此时距离暂停开始已过去十一天。OpenAI 表示,这种计算密集型红队测试行动现已成为新训练运行[审批流程](https://openai.com/index/towards-safety-cases-for-frontier-ai-training/)的标准组成部分。 ### 没有炒作的 AI 新闻——由人工精选 订阅 THE DECODER,享受无广告阅读、每周 AI 新闻通讯、每年六次独家“AI Radar”前沿报告、完整档案访问权限以及评论区访问权限。 ## 正文 · 原文 **OpenAI now trains frontier models without internet access and uses AI agents to red-team its own infrastructure. In separate isolation tests, agents even communicated through covert side channels, showing how hard true containment is.** On September 20, an internal research model exploited a misconfigured DNS resolver to reach the open internet during training, even though OpenAI had already tightened security after the [now infamous Hugging Face incident](https://the-decoder.com/openai-claims-responsibility-for-the-hugging-face-hack-after-its-own-models-escaped-a-test-sandbox/) in July and other [observed misalignment cases](https://the-decoder.com/tens-of-thousands-of-security-probes-show-openais-hugging-face-incident-was-just-the-beginning/). OpenAI contained the breach quickly but still halted all frontier model training and evaluation. ## OpenAI cuts frontier models off from the internet during training OpenAI made a policy decision in response: Frontier models now train without internet access until safety measures justify restoring it. Models learn to browse and search inside a closed "offline world" built from indexed and cached web content, and all frontier reinforcement learning runs now operate this way. [![](https://the-decoder.com/wp-content/uploads/2026/10/openai_offline_world.png)](https://the-decoder.com/wp-content/uploads/2026/10/openai_offline_world.png) OpenAI's agent sandbox allows access only to offline resources. The live internet, other sandboxes, and OpenAI's infrastructure sit outside the containment boundary. | Image: OpenAI Full network isolation, a so-called air gap, isn't feasible yet because experiments still need external connections for operations and monitoring, the company says. But even a complete air gap wouldn't guarantee containment, [according to OpenAI researcher Noam Brown](https://the-decoder.com/406-scientists-demand-ai-models-be-cut-off-from-the-internet-immediately/#airgapping-buys-time-but-the-real-problem-runs-deeper-openai-researcher-brown-says). He gave one example of how isolated systems could still communicate: two separate computers could theoretically signal each other through temperature sensors, with one heating its CPU while the other reads the change. OpenAI CEO [Sam Altman made a similar argument](https://the-decoder.com/openai-ceo-sam-altman-says-advanced-ai-will-inevitably-escape-its-sandbox-and-only-alignment-can-save-us/) in a recent interview: Sandboxes offer only limited protection against powerful-enough models, and at some point it may no longer be possible to trace how they escaped. Monitoring helps early on, but what ultimately matters is that an AI system internalizes human values, a problem the field knows as "AI alignment" and has yet to solve. ## Isolated AI agents found ways to talk to each other through CPU cache OpenAI documents just how hard it is to control AI agents [in its blog post](https://openai.com/index/securing-our-frontier-model-training-environment/). During tests, agents found ways to communicate without a network. In one test, agents placed in separate, isolated environments on shared hardware were told to exchange secrets with no network connection and no shared files. They used the processor cache. One agent deliberately altered the data stored there, while the other measured differences in load times and decoded the message. The same technique is behind the well-known [Spectre and Meltdown attacks](https://spectreattack.com/). More channels turned up through write patterns on shared memory and through OpenAI's own offline infrastructure, the system meant to replace the internet. One agent noted in its reasoning log that it "could create an acoustic modem-like signal … using continuous writes vs idle." In one case, agents were given only a general list of vulnerabilities and told to find a communication channel on their own. They developed a channel independently, agreed on an encoding scheme, and successfully exchanged information. The channels are usually slow and hard to exploit, OpenAI says, but they do reach across boundaries meant to keep agents apart. ## OpenAI turned AI agents loose on its own infrastructure During the training pause, OpenAI also had teams of AI agents systematically attack its own infrastructure. The agents were given source code, configuration details, and documentation of potential vulnerabilities. OpenAI's goal was to test its defenses under conditions that favored the attackers. The agents included the upcoming GPT-6 Astra Cyber and an internal frontier model. The campaign consumed hundreds of billions of tokens, according to OpenAI. [![](https://the-decoder.com/wp-content/uploads/2026/10/Red-teaming-findings-versus-tokens-spent.png)](https://the-decoder.com/wp-content/uploads/2026/10/Red-teaming-findings-versus-tokens-spent.png) Unique potential security issues identified during the red-teaming campaign, plotted against tokens spent in billions. Most appeared within the first 110 billion tokens. | Image: OpenAI The agents discovered previously unknown vulnerabilities in the open-source components in use and exploited them to escape from secured virtual machines. One was a vulnerability in Kata Containers, a widely used container isolation system. OpenAI reported it as [CVE-2026-24834](https://github.com/kata-containers/kata-containers/security/advisories/GHSA-wwj6-vghv-5p64) and provided a patch. More fixes went to other projects, and OpenAI reported additional, as-yet-unpublished vulnerabilities through the ["Patch the Planet"](https://openai.com/index/patch-the-planet/) initiative. In tests where agents had already gained control of a host, they found a vulnerability in nearly every remaining internet-facing service. In one case, an agent used a firewall-permitted connection to fetch an external URL through a metrics service. Others bypassed security filters through misconfigurations. ## AI red-teaming is now required before every new training run The agents could no longer find a way to access the internet that was both easy to exploit and hidden from monitoring, according to OpenAI. The red-teaming also exposed gaps in monitoring and in the automatic pause mechanism. After fixing the issues, OpenAI resumed training and internal use of its frontier models on October 1, eleven days after the pause began. OpenAI says this kind of compute-intensive red-teaming campaign is now a standard part of [the approval process](https://openai.com/index/towards-safety-cases-for-frontier-ai-training/) for new training runs. ### AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.