Nvidia 推出平台，可在“毫秒级”隔离失控 AI Agent

发布于 28/09/2026 - 12:46 GMT+2

Nvidia 正在推出一个面向人工智能（AI）Agent 的安全平台，已有超过 100 家机构正与其合作，包括 Anthropic、SpaceXAI 和 JPMorganChase。

该平台名为 Nvidia Open Agent Safety Platform。根据这家美国芯片制造商发布的公告，它旨在为 Agent 可访问的资源设置限制；一旦 Agent 越过这些限制，平台可在“毫秒级”将其置于“隔离”状态。

该平台可用于运行 Agent 的软件与硬件、算力及机器人系统。

Nvidia 创始人兼 CEO 黄仁勋（Jensen Huang）表示：

“只有解决 AI 安全问题，AI 为社会带来的非凡潜力才能真正实现。”

“随着我们持续探索 AI 能力的前沿，必须加速探索 AI 安全的前沿……携手合作，我们能够提升全球 AI 安全的标准。”

Nvidia Open Agent Safety Platform 由 OpenShell 和 Sentry 两项工具组成。

OpenShell 运行于中央处理器（CPU）上，CPU 负责处理运行 AI Agent 所需的许多任务。Nvidia 表示，该工具可在其最新、专为 AI Agent 设计的 Vera CPU 上高效运行，也可适配其他公司生产的芯片。

Sentry 则是一套额外的“看门狗”系统，可运行在独立的 Nvidia 芯片上，用于监测绕过软件控制措施的 Agent。

Nvidia 表示，参与合作的机构正在以不同方式使用或开发这项技术。

Anthropic 已与 Nvidia 合作，为其企业用于运行 AI Agent 的 Claude 服务添加安全控制功能，使企业能够限制这些 Agent 可访问的资源。

SpaceX 的 AI 部门 SpaceXAI 表示，其正将该平台用于 Cursor 编程 Agent 和 Grok 模型；Salesforce 则已将 OpenShell 与 Slack 集成，使团队能够查看 Agent 的操作，并批准或拒绝其提出的额外权限请求。

在金融领域，JPMorganChase 和 Citi 正与 Nvidia 合作开发共享的开源 Agent 安全技术。

AI 公司呼吁加强防护措施

此次公告发布之际，业界对加强 AI Agent 防护措施的呼声正不断升高。

7 月，OpenAI 承认，其一款模型利用隐藏漏洞逃离受控测试环境，并入侵了 Hugging Face 的服务器。

与此同时，Anthropic 表示，其 AI 模型在测试阶段入侵了三家机构的系统。

Nvidia 表示，这些事件表明，Agent 在尝试完成被分配任务时，可能绕过应用程序内置的安全控制；而“安全与安保需要全栈工程能力”。

Published on 28/09/2026 - 12:46 GMT+2

Nvidia is rolling out a security platform for artificial intelligence (AI) agents that more than 100 organisations are working with, including Anthropic, SpaceXAI and JPMorganChase.

ADVERTISEMENT

ADVERTISEMENT

Called the Nvidia Open Agent Safety Platform, it is designed to set limits on what agents can access and put them in “quarantine” within milliseconds if they cross those limits, according to an announcement from the US chipmaker.

The platform can be used across software and hardware, compute and robotics systems that run agents.

“AI’s extraordinary potential for society will only be realised if we solve AI safety,” said Jensen Huang, the CEO and founder of Nvidia.

“As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety…. Together, we can raise the bar for global AI safety.”

The Nvidia Open Agent Safety Platform consists of two tools, OpenShell and Sentry.

OpenShell runs on central processing units (CPUs), the chips that handle many of the tasks needed to run AI agents. Nvidia says it works efficiently on Vera, its latest CPU designed for AI agents, and can also be adapted for chips made by other companies.

Sentry is an additional "watchdog" system that would run on a separate Nvidia chip, watching for agents that get past the software’s controls.

The organisations working with Nvidia are using or developing the technology in different ways, Nvidia says.

Anthropic has worked with Nvidia to add security controls to its Claude service for businesses to run AI agents, allowing companies to restrict what those agents can access.

SpaceX’s AI division, SpaceXAI, says it is using the platform for Cursor coding agents and Grok models, while Salesforce has integrated OpenShell with Slack so teams can see what agents are doing and approve or reject requests for additional permissions.

In finance, JPMorganChase and Citi are collaborating with Nvidia on shared open-source agent safety technology.

AI companies call for stronger safeguards

The announcement comes amid growing calls for stronger safeguards around AI agents.

In July, OpenAI admitted one of its models exploited a hidden flaw to escape a controlled test and break into Hugging Face's servers.

Meanwhile, Anthropic said its AI model hacked into three organisations' systems during the testing phase.

Nvidia said the incidents showed that agents could get around security controls built into an app while trying to complete their assigned tasks and that “safety and security require full-stack engineering”.