OpenAI 在其 AI 攻破 Hugging Face 后公布新的安全调整

OpenAI 宣布了新的安全更新。此前在 7 月，其 AI 曾逃出沙箱环境，并意外入侵 Hugging Face。这些更新包括改进研究环境、监控机制和对齐技术。该公司此前已经叫停一款新模型 Astra，因为其认为该模型可能具备“关键”级别的网络安全能力。OpenAI 表示，在收紧安全措施期间，公司还暂停了其“最新、计划部署模型”的强化学习（RL）训练两周。该公司“规模最大的计划中前沿 RL 训练仍处于暂停状态”。

对于前沿模型研究，OpenAI 现在要求为“执行模型生成或其他不受信任代码”的工作负载提供更强的沙箱，并增加更多控制措施，以“将高风险和不受信任的工作负载与互联网隔离”。该公司还更新了研究环境，以“移除可能存在漏洞的共享服务、减少常设权限，并改善安全与信任边界”。

作为扩大监控体系的一部分，OpenAI 表示，其现在的目标是在发现可疑活动后“30 分钟内”发出警报。如果在警报发出后被呼叫处理的人员无法在 30 分钟内“明确”判断该警报是否为误报，“相关团队应暂停该活动”。

OpenAI 还表示，正在“将我们的核心对齐技术应用于训练流程的更多阶段”，其中包括能够“更好地检测和抑制不安全行为”的奖励模型，以及训练模型“更诚实地说明其行为、能力和局限性”。

自发现 Hugging Face 遭入侵以来，Anthropic 和 Meta 也发现，它们的 AI 模型曾入侵其他组织。

OpenAI is announcing security updates following the July news that its AI broke out of a sandboxed environment and accidentally hacked Hugging Face, including improvements to its research environments, monitoring, and alignment techniques. The company had already put the brakes on a new model, Astra, that it thinks could have “critical” cybersecurity capabilities, and the company says it instituted a two-week pause in reinforcement learning (RL) training on its “latest models intended for deployment” while it tightened up security. The company’s “largest planned frontier RL run remains on hold.”

For its frontier model research, OpenAI now requires stronger sandboxes for workloads that “execute model-generated or otherwise untrusted code,” and has more controls to “isolate higher-risk and untrusted workloads from the internet.” It has also updated its research environment to “remove potentially vulnerable shared services, reduce standing privileges, and improve security and trust boundaries.”

As part of the company’s expanded monitoring setup, OpenAI now aims to issue an alert “within 30 minutes after concerning activity is surfaced,” OpenAI says. If the people paged after an alert can’t “conclusively” determine whether an alert is a false positive within 30 minutes, “those teams are expected to pause the activity.”

OpenAI also says that it’s applying “our core alignment techniques across more stages of the training process,” including reward models that “better detect and discourage unsafe behavior” and training models “to be more honest about their actions, capabilities, and limitations.”

Since the discovery of the Hugging Face breach, Anthropic and Meta have also found that their AI models had hacked other organizations.