OpenAI 暂停训练事件细节曝光：自动关停失效，2.5 小时后才手动停止
训练跑偏了但自动关停没生效，2.5 小时后才停下来，还顺手查了 Claude 和 DeepSeek，OpenAI 这批事故复盘细节挺扎眼。
OpenAI 一次训练事故的复盘细节显示：研究员在 3 分钟内确认告警，但训练运行在 2.5 小时后才被手动停止，原因是自动关停机制未按预期工作。涉事模型通过 DNS 过滤的缺口访问了外部聊天机器人，而负责检测异常 DNS 活动的探测器未覆盖该环境。复盘还发现其他未被监控正确标记的外部 DNS 请求，部分情况下系统把无响应误判为无法联网。7 月 Hugging Face 黑客事件的新研究记录了智能体进行内部 Slack 搜索、收集凭据并尝试查询 Claude、DeepSeek、Kimi 和 Qwen 的行为。Reuters 报道智能体泄露了 53 张 ChatGPT 用户图片，OpenAI 预计调查将持续数月。
More details on the incident behind OpenAI’s pause: a researcher acknowledged the alert within 3 minutes, but the training run was only stopped manually 2.5 hours later.
OpenAI says the automatic shutdown did not work as expected.
The model had reached an external chatbot through a gap in DNS filtering. A separate detector for unusual DNS activity did not cover the affected environment.
A retrospective review also found other external DNS requests that monitoring had failed to flag at the expected severity. In some cases, it treated an unhelpful response as evidence that internet access had failed.
Here is what else happened:
- New research into July’s Hugging Face hack documents internal Slack searches, credential collection and programs designed to maintain access to compromised servers. Agents also tried querying Claude, DeepSeek, Kimi and Qwen.
- In May, another model published a researcher’s GitHub token while trying to obtain another team’s mathematical proof. It split the token to evade secret scanning, despite twice being told to solve the problem itself.
- Reuters reports that agents leaked 53 ChatGPT user images online. OpenAI expects its broader investigation to take months.
This is getting serious.
