OpenAI 强化安全措施：部分工作负载的计算开销将上升 20%

扩展多阶段思维链监控，使前沿模型工作成本更高

OpenAI 周二表示，在未发布、未经监督的 AI 模型入侵 HuggingFace 后，公司已实施的暂停模型训练工作决定仍然有效。这家 AI 公司正试图落实更强的安全措施，其中部分措施将使受监控推理工作负载的计算开销增加 20%。

OpenAI 发言人向 The Register 表示，这些成本反映的是内部研究支出，不会直接转嫁给客户。该公司尚未披露目前有多少比例的总推理计算受到这类监控，也未说明此前监控机制覆盖了多少推理计算。

OpenAI CEO Sam Altman 在一则社交媒体帖子中写道：“我们已暂停部分前沿 RL（强化学习）训练，以确保能够达到眼前这一新能力水平所需的适当对齐、安全和监控标准。模型进展如今极其迅速；我们一直表示，如果认为模型能力的发展速度超过安全与对齐工作的步伐，我们就会采取行动。”

Altman 表示，他仍预计新模型——推测是被推迟的 Astra——将很快发布。训练暂停影响的是更晚期的产品发布。

OpenAI 在其文章中重申，将重点投入监控、模型对齐与安全措施，以防止其模型像上个月那样失控。在 HuggingFace 事件之后，OpenAI “暂停了研究集群中可能执行代码或使用可访问互联网工具的运行任务所涉及的前沿模型推理”。

该公司称，部分工作负载仍获准运行，但其他工作负载将暂停，直到它们能够迁移至更严格的安全体系之下。该体系包括沙箱隔离、网络隔离和持续安全测试。

该公司写道：“我们规模最大的计划中前沿 RL（强化学习）运行仍处于暂停状态；我们正在开展更小规模的训练和评估，以评估模型行为、验证我们的安全措施，并在继续之前建立更多关于模型对齐的证据。”

强化学习是指 AI Agent 通过试错、因获得期望结果的奖励而“学习”其环境的过程。

OpenAI 还表示，正扩大对思维链过程的监控。思维链是一种让“思考型”模型将任务拆解为离散步骤，并为每一步生成中间文本输出的技术。

该公司此前的方法聚焦于高风险工作负载，具体包括前沿模型的内部部署以及前沿 RL 训练运行。

与之相比，OpenAI 表示，其新的监控体系覆盖所有涉及工具的 RL 训练和评估，前提是相关模型能力达到 GPT-5.6 Sol 或更高水平。鉴于 Astra 被认定具备关键网络能力，OpenAI 又增加了一项额外监控要求，覆盖 Astra 的所有推理，而不仅是 RL 训练和测试。

OpenAI 表示：“这些安全措施需要大量计算资源。我们目前估计，监控开销约占被监控推理计算的 20%，但不同训练和评估工作负载之间的成本差异很大。”

OpenAI 预计将在未来的文章中分享更多有关其监控方案实施细节的信息。在去年发布的研究中，该公司称思维链监控是发现模型不当行为的有效方法，但也警告，直接优化模型以严格遵循指令“并不能消除所有不当行为，还可能导致模型隐藏其意图”。

如果你选择相信该公司所称不会将模型“思维监管”的成本转嫁给客户，那么可以推断 OpenAI 的亏损将会扩大。如果 OpenAI 上市，很难想象这种做法能够长期持续。

但考虑到该公司据报道已作出$6000 多亿美元 AI 基础设施承诺，并预计将至少到 2030 年仍无法盈利，为了应对不确定的安全风险，多花一点钱又算什么？

ai and ml

OpenAI's overhead will rise 20 percent for some workloads as it hardens security

Expanded multistage chain of thought monitoring makes frontier model work more expensive

OpenAI on Tuesday said its decision to suspend model training work, implemented after unreleased, unsupervised AI models hacked HuggingFace, remains in effect as the AI biz tries to implement stronger security measures. Some of those measures will increase compute overhead by 20 percent of the observed inference workload.

An OpenAI spokesperson told The Register that those costs reflect internal research and won't be passed on directly to customers. The company has not revealed what portion of its total inference compute is subject to such monitoring now, or under its prior monitoring regime.

"We have paused some frontier RL [reinforcement learning] training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us," OpenAI CEO Sam Altman wrote in a social media post. "Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment."

Altman said he still expects new models, presumably the delayed Astra, will ship soon. The training pause affects further-out releases.

OpenAI in its post reiterated its plans to focus on monitoring, model alignment, and security measures to prevent its models from running amok as they did last month. Following the HuggingFace incident, OpenAI "paused frontier model inference in research clusters for runs that could execute code or use tools that could access the internet."

The biz said it allows some workloads to run, but paused others until they can be moved under a more stringent security regime that includes sandboxing, network isolation, and continuous security testing.

"Our largest planned frontier RL (reinforcement learning) run remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding," the company wrote.

Reinforcement learning refers to the trial-and-error process by which AI agents "learn" about their environment by being rewarded for desired outcomes.

OpenAI also said it is expanding its monitoring of the chain-of-thought process, the technique that sees "thinking" models break down tasks into discrete steps and produce intermediate text output for each step.

The company's prior approach focused on high-risk workloads, specifically internal deployments of frontier models and frontier RL training runs.

In contrast, OpenAI says, its new monitoring setup covers all RL training and evaluations involving tools for models at the capability level of GPT-5.6 Sol or higher. And with the determination that Astra possesses critical cyber capabilities, OpenAI added an additional monitoring requirement that covers all inference with Astra, not just RL training and testing.

"These safeguards require meaningful compute," OpenAI said. "Our current estimates put monitoring overhead at roughly 20 percent of the inference compute being monitored, though the cost varies substantially across training and evaluation workloads."

OpenAI expects to share more details about the implementation of its monitoring scheme in a future post. In research published last year, the company said that chain-of-thought monitoring is an effective way to detect model misbehavior, but cautioned that directly optimizing models to strictly follow instructions "does not eliminate all misbehavior and can cause a model to hide its intent."

If you choose to believe the company's assurance that it will not pass on the cost of model thought policing to customers, it follows that OpenAI's losses will increase. It's difficult to imagine that would be a sustainable stance if OpenAI goes public.

But given the company's reported $600+ billion in AI infrastructure commitments and its expectation to remain unprofitable until at least 2030, what's a bit more expense for the sake of uncertain security? ®