Nvidia 希望在每个 AI Agent 旁部署一枚“看门狗”芯片

Nvidia 正在推出一个新的软件平台，让 AI 开发者能够为 Agent 设置安全护栏，并防止它们逃离隔离环境。

Nvidia CEO 黄仁勋周一在 CNBC 节目《Squawk Box》中表示：“你不能让 Agent 在公司内部四处游荡、不断漂移，因此必须找到一种方法把它们隔离起来。”

黄仁勋表示，这一新平台本质上是“面向 Agent 的浏览器”，提供了一套隔离系统，只允许 Agent 访问完成其工作所需的资源。

Nvidia 周一发布 Open Agent Safety Platform，此前包括 OpenAI、Anthropic、Meta 和 Google 在内的公司都曾披露，其人工智能模型近期发生了事件：模型逃出沙箱，并试图入侵其他公司、访问其计算机系统。

Nvidia 一名代表周日在电话会议上对记者表示，其平台本可以阻止 OpenAI 今年 7 月涉及 Hugging Face 的事件。当时，OpenAI 的模型逃离了隔离环境，访问开放互联网并入侵了运营开源开发者平台的 Hugging Face。

“每一起安全事件都具有独特性，我们必须逐一详细审视，”全球市值最高公司 Nvidia 的企业 AI 副总裁 Justin Boitano 表示。“据我们所知，Hugging Face 报告称，超过 1.7 万个 Agent 对其基础设施发动攻击，持续了数天乃至数周。”

自近四年前 ChatGPT 推出以来，Nvidia 一直处于生成式 AI 热潮的核心。该芯片制造商的图形处理器对于大语言模型开发，以及超大规模云服务商提供的 AI 服务都至关重要。不过，黄仁勋近来已成为 AI 安全辩论中的重要声音。他认为，许多安全担忧属于工程问题，可以通过计算机科学和产品开发加以解决。

黄仁勋上周在接受《纽约时报》Ezra Klein 播客采访时，针对近期事件表示：“你必须思考自己原本可以做什么，解决方案是什么。未来要改进流程，以避免这种事情再次发生。”

两周前，Anthropic CEO Dario Amodei 呼吁 AI 模型开发者放缓能力提升速度，原因是担忧模型失去控制。这一观点得到了 OpenAI 的 Sam Altman 和 SpaceX 的 Elon Musk 支持，并在行业内引发激烈争论。

Boitano 表示，Nvidia 的新产品是针对 Agent 安全问题的一项工程解决方案。

Boitano 表示：“近期事件凸显了 AI Agent 面临的一项根本障碍：仅靠模型层面的安全护栏，无法治理 Agent 可以访问什么或能够做什么。”

该平台的一个组件名为 Nvidia OpenShell，运行于中央处理器之上，并为 Agent 的能力设定限制。Nvidia 还发布了 Sentry，它负责监控 Agent，运行在网络芯片上，而非 CPU 或 GPU 上。

部分软件采用开源形式。Nvidia 将该平台称为参考设计，这意味着合作伙伴将基于它构建产品并推向市场。

Nvidia 指定 Cisco、Microsoft、Oracle、CoreWeave、Dell、HPE、Lenovo、ARM 和 Intel 为合作伙伴。Nvidia 还正与 Anthropic 合作，将云端托管 Agent 与 OpenShell 集成。

黄仁勋周一对 CNBC 表示：“如果世界不认为这个行业是以安全方式构建的，或者不确信它是以安全方式构建和部署的，我们就不可能拥有一个成功的 AI 行业。”

Nvidia is rolling out a new software platform to allow AI developers to set safeguards for agents and prevent them from breaking out of containment.

"You can't have agents roam around and drift around the company, and so you have to find a way to container it," Nvidia CEO Jensen Huang told CNBC's "Squawk Box" on Monday.

Huang said the new platform is essentially "a browser for agents," providing a containment system that only allows access to things an agent needs to do its job.

The release on Monday of Nvidia's Open Agent Safety Platform comes after companies including OpenAI, Anthropic, Meta, and Google disclosed recent incidents in which their artificial intelligence models escaped their sandboxes and attempted to hack other companies and access their computer systems.

An Nvidia representative told reporters on a call on Sunday that its platform could have prevented OpenAI's Hugging Face incident in July. That's when OpenAI models escaped containment, accessed the open internet and breached Hugging Face, which operates an open-source developer platform.

"Each security incident is unique, and we have to look at all of them in detail," said Justin Boitano, vice president of enterprise AI at Nvidia, the world's most valuable company. "From what we know, Hugging Face reported over 17,000 agents attacking their infrastructure that went on for days and weeks."

Nvidia has been at the center of the generative AI boom since the launch of ChatGPT almost four years ago, as the chipmaker's graphics processing units are critical to the development of large language models and to the AI services offered by hyperscalers. But Huang has more recently emerged as a key voice in the AI safety debate, arguing that many security concerns are engineering issues that can be solved through computer science and product development.

"You have to think about what you could have done, what's the solution for it," Huang said in a podcast with The New York Times' Ezra Klein released last week, referring to recent incidents. "In the future, improve your process so that you could avoid this from happening again."

Anthropic CEO Dario Amodei set off an industry firestorm two weeks ago, urging AI model developers to slow their pace of advancement due to fears of the models spinning out of control, an argument that was supported by OpenAI's Sam Altman and SpaceX's Elon Musk.

Nvidia's new offering is an engineering solution to the agent safety issue, Boitano said.

"Recent incidents have highlighted a fundamental hurdle for AI agents, and that is that model-level safeguards alone can't govern what agents can access or do," Boitano said.

One component of the platform is called Nvidia OpenShell, which runs on central processors and sets limits on agent capabilities. Nvidia also announced Sentry, which monitors agents and runs on network chips, not CPUs or GPUs.

Some of the software is open source, and Nvidia is calling its platform a reference design, which means partners are intended to build products on top of it to bring it to market.

Nvidia named Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM and Intel as partners. Nvidia is also working with Anthropic to integrate cloud managed agents with OpenShell.

"We can't have a successful AI industry if the world doesn't think it's built or confident that it's built and deployed safely," Huang told CNBC on Monday.