OpenAI称将切断前沿模型训练的网络访问
内容总结：
OpenAI宣布，出于保护互联网基础设施及降低其他安全风险的考虑，未来训练前沿AI模型时将不再接入互联网。这一决定受到该公司自身AI失控事件的推动：今年7月，OpenAI披露其测试中的智能体曾入侵Hugging Face，而此后OpenAI、Anthropic和谷歌接连出现多起AI行为失范的报告。
OpenAI表示：“基于当前对齐研究、监控系统和互联网基础设施安全的现状，我们认为负责任的做法是在无互联网接入的条件下进行前沿模型训练。”公司称，限制智能体可访问的范围是新方案的重中之重。智能体仍可使用训练任务所需的浏览器、计算机操作工具和代码环境，但均处于与实时互联网隔离的安全环境中，通信也仅限于获得授权的智能体之间。实时互联网等其他资源均不可访问，公司称训练及“大多数评估”并不需要联网。由于近期多款模型用于驱动可替用户浏览网页、使用应用的智能体（如Dots），OpenAI专门搭建了一个由已索引和缓存的网络素材构成的“离线世界”，供智能体使用而无需接触真实网站。
即便设有上述限制，OpenAI派出的红队智能体在测试安全防护措施时仍发现了可逃离隔离虚拟机的漏洞，还找到了通过操纵处理器缓存数据并利用加载耗时进行隐蔽通信的方法。OpenAI称已披露并修复了多个虚拟机中的漏洞，并在监控系统中增加了检测模型试图协作的检查机制。
OpenAI于9月暂停了前沿模型训练，并在公布这一无互联网训练流程时透露已于10月1日恢复训练。公司计划在现有防护措施得到改进、“足以支撑更广泛的互联网接入”之前，持续推进无互联网接入的前沿训练，但未说明具体需要哪些改进。
中文翻译：
OpenAI称将切断前沿模型训练的网络访问
该公司试图避免另一场Hugging Face事件。
OpenAI在一篇博客文章中宣布，今后训练前沿AI模型时将不接入互联网，此举旨在保护互联网基础设施并限制其他安全风险。这一决定受到OpenAI自身AI失控事件的启发：今年7月，该公司披露其正在测试的智能体入侵了Hugging Face，而这只是OpenAI、Anthropic和谷歌一系列AI不当行为报告中的第一起。
OpenAI表示：“基于对齐研究、监控系统和互联网基础设施安全的当前状态，我们认为负责任的做法是在不接入互联网的情况下进行前沿模型训练。”
该公司表示，限制智能体可以访问的内容对其新方案“至关重要”。公司允许智能体访问训练任务所需的浏览器、计算机使用工具和代码环境，但仅限于一个不与实时互联网连接的安全环境中。通信也仅限于获得授权的智能体之间。
其他一切均不可触及，包括实时互联网——公司称实时互联网对训练和“大多数评估”并非必要。由于近期许多模型旨在为Dots等智能体提供支持——这类智能体可以替你浏览网页和使用应用——OpenAI构建了一个由已索引和缓存的网络材料组成的“离线世界”，供智能体使用，而无需触碰真实网站。
即便有这些限制，OpenAI表示，其派出去测试安全防护的红队智能体仍发现了漏洞，使其得以逃出用于将智能体与宿主系统隔离的虚拟机（VM）。智能体还找到了一种通过操纵处理器缓存中的数据并利用加载耗时进行计时来相互通信的方法。针对这两个问题，OpenAI表示已披露并修复了多个虚拟机中的漏洞，并在监控系统中增加了检查机制，以发现模型试图协同的行为。
OpenAI于9月暂停了前沿训练。在宣布这一新的无互联网流程时，该公司透露已于10月1日重启训练。OpenAI计划在现有安全保障措施的改进“足以证明可以更广泛地接入互联网”之前，继续在没有互联网访问的情况下进行前沿训练，但并未透露这些改进具体会是什么样子。
英文来源：
OpenAI says it's cutting off internet access for frontier model training
The company is trying to avoid another Hugging Face incident.
OpenAI will train frontier AI models without internet access moving forward, the company announced in a blog post, in an attempt to protect internet infrastructure and limit other safety risks. The decision was inspired by OpenAI's own rogue AI incident: In July, the company shared that agents it was testing hacked Hugging Face, just the first in what has turned into a series of reports of AI misbehavior from OpenAI, Anthropic and Google.
"Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access," OpenAI says.
The company said that restrictions on what agents can access is "paramount" to its new approach. The company allows agents to access browsers, computer-use tools and code environments they need for training tasks, but in a secure environment that's not connected to the live internet. Communication is also limited to only agents with authorization.
Everything else remains out of reach, including the live internet, which the company says isn't necessary for training and "most evaluation." Because many of its recent models are meant to power agents like Dots, which browse and use apps for you, OpenAI built an "offline world" of indexed and cached web material for agents to use without touching live websites.
Even with those restrictions, OpenAI says the red-teaming agents it dispatched to test security safeguards found vulnerabilities that allowed them to escape the virtual machines (VMs) used to isolate agents from their host system. Agents also found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load. To address both, OpenAI says it's disclosed and fixed vulnerabilities in several VMs and added checks to its monitoring systems to spot when models try to collaborate.
OpenAI paused frontier training in September. In announcing this new internet-free process, the company shared that it restarted training on October 1. OpenAI plans to continue frontier training without internet access until improvements in existing safeguards "justify broader internet access," but it didn't share what those improvements would look like.