原帖引用 · AI 翻译
在 Hugging Face 遭入侵后，OpenAI 将不对齐行为监控扩展至前沿模型的训练和部署过程。
OpenAI 披露了用于发现前沿 AI 智能体不当行为的 4 层监控系统。
OpenAI 表示，Hugging Face 遭入侵事件始于其内部网络安全评估，当时 OpenAI 没有运行不对齐行为监控。
> 此前，OpenAI 的不对齐行为监控主要针对其内部编程智能体，因此 Hugging Face 遭入侵事件始于无人监控不当行为的网络安全评估。
> 现在，监控覆盖 GPT-6 Astra 级模型的每一次强化学习运行、评估和工具使用配置、所有内部 Codex 流量，以及客户对 Astra 级网络安全模型的工具使用。
> 活动最多经过 4 层筛查，从低成本的激活探针，到 AI 审核员，再到检查网络日志的调查智能体；最高级别的警报通常会在 15 分钟内呼叫人工处理。
> OpenAI 表示，如果将这套系统用于重放 Hugging Face 事件，安全团队本应在沙箱逃逸约 30 分钟后、Hugging Face 遭入侵 30 多小时前收到警报。
原文
After the Hugging Face breach, OpenAI extended misalignment monitoring across frontier model training and deployment. OpenAI has revealed the 4-layer monitoring system it uses to catch frontier AI agents misbehaving. Says the Hugging Face breach began in internal cybersecurity evaluations where OpenAI ran no misalignment monitors. > Before, OpenAI’s misalignment monitors mainly watched its internal coding agents, so the Hugging Face breach began in cyber evaluations that nobody was monitoring for misbehavior. > Now the monitors cover every RL run, evaluation and tool-use setup for GPT-6 Astra-class models, all internal Codex traffic, and customer tool use of Astra-class cyber models. > Activity is screened in up to 4 layers, from cheap activation probes to an AI reviewer to investigator agents checking network logs, and top alerts typically page humans within 15 minutes. > OpenAI says this setup, replayed on the Hugging Face incident, would have alerted security about 30 minutes after the sandbox escape and more than 30 hours before Hugging Face was breached.