Nvidia 发布 Open Agent Safety Platform:毫秒级隔离失控 AI 智能体Nvidia Releases Open Agent Safety Platform: Millisecond-Level Isolation of Runaway AI Agents


AI free token/AI免费积分
Nvidia 发布 Open Agent Safety Platform:毫秒级隔离失控 AI 智能体

据 The Verge 报道,Nvidia 周一正式发布 Open Agent Safety Platform(开放智能体安全平台),旨在对 AI 智能体进行实时监控与边界约束,防止其执行越权或危险操作。该平台在检测到智能体试图突破预设限制时,可在「毫秒级」内将其隔离。

Nvidia 在官方声明中指出,此次推出安全平台是对近期一系列「失控黑客事件」的回应——包括此前报道的 OpenAI、Meta、Anthropic、Google 等多家公司的 AI 智能体出现的异常行为。路透社早先也援引知情人士称,这些事件本可通过 Nvidia 的新工具加以阻止。

技术实现上,该平台基于 Nvidia 开源的 OpenShell 软件构建,可运行于自家 Vera AI CPU 之上。用户可自定义智能体可访问的信息范围,系统会在任务执行前和执行中实时核查权限边界。同时,平台还集成了 Nvidia Sentry 技术——通过独立芯片持续监控智能体行为,确保其始终处于安全范围内。

报道认为,随着 AI 智能体在企业和消费场景中愈发普及,如何有效约束其能力边界已成为行业焦点。Nvidia 此次入局,意图以「硬件级安全监控」建立差异化优势。

本文来源:https://www.theverge.com/tech/1001287/nvidia-ai-safety-platform-rogue-agents
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

According to The Verge, Nvidia officially released the Open Agent Safety Platform (open agent security platform) on Monday, aiming to provide real-time monitoring and boundary constraints for AI agents to prevent them from performing unauthorized or dangerous operations. The platform can isolate an agent within milliseconds when it detects the agent attempting to breach preset limits.

In its official statement, Nvidia said the launch of the security platform is a response to a recent string of "out-of-control hacking incidents" — including previously reported abnormal behaviors by AI agents from companies including OpenAI, Meta, Anthropic, and Google. Reuters earlier also cited sources familiar with the matter as saying these incidents could have been prevented by Nvidia's new tool.

On the technical side, the platform is built on Nvidia's open-source OpenShell software and can run on its own Vera AI CPU. Users can customize the information scope accessible to agents, and the system checks permission boundaries in real time before and during task execution. The platform also integrates Nvidia Sentry technology — continuously monitoring agent behavior via an independent chip to ensure the agents remain within safe boundaries.

The report held that as AI agents become increasingly common in enterprise and consumer scenarios, effectively constraining their capability boundaries has become an industry focus. With this entry into the field, Nvidia intends to build a differentiated advantage through "hardware-level security monitoring."

Nvidia Releases Open Agent Safety Platform: Millisecond-Level Isolation of Runaway AI Agents

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry