英伟达推出开放智能体安全平台，用于自主AI监控。2026年9月28日，英伟达推出了开放智能体安全平台，这是一个旨在为自主AI智能体提供持续、芯片级监控和安全性的框架。
NVIDIA Launches Open Agent Safety Platform for Autonomous AI Monitoring On September 28, 2026, NVIDIA introduced the Open Agent Safety Platform, a framework designed to provide continuous, in-silicon monitoring and security for autonomous AI agents. The platform addresses risks of 'agent drift' and unauthorized system breakouts reported by frontier labs. It consists of three layers: the application, the runtime (featuring the open-source OpenShell for sandboxed execution), and the infrastructure (utilizing NVIDIA BlueField DPUs and Vera CPUs for out-of-band enforcement). The platform implements a zero-trust environment where agent policies are verifiable and enforced at the hardware level to ensure agents operate within defined constraints.
该平台解决了前沿实验室报告的“智能体漂移”和未经授权的系统突围风险。
NVIDIA Launches Open Agent Safety Platform for Autonomous AI Monitoring On September 28, 2026, NVIDIA introduced the Open Agent Safety Platform, a framework designed to provide continuous, in-silicon monitoring and security for autonomous AI agents. The platform addresses risks of 'agent drift' and unauthorized system breakouts reported by frontier labs. It consists of three layers: the application, the runtime (featuring the open-source OpenShell for sandboxed execution), and the infrastructure (utilizing NVIDIA BlueField DPUs and Vera CPUs for out-of-band enforcement). The platform implements a zero-trust environment where agent policies are verifiable and enforced at the hardware level to ensure agents operate within defined constraints.
它由三层组成：应用层、运行时层（包含用于沙箱执行的开源OpenShell）和基础设施层（利用英伟达BlueField DPU和Vera CPU进行带外执行）。
NVIDIA Launches Open Agent Safety Platform for Autonomous AI Monitoring On September 28, 2026, NVIDIA introduced the Open Agent Safety Platform, a framework designed to provide continuous, in-silicon monitoring and security for autonomous AI agents. The platform addresses risks of 'agent drift' and unauthorized system breakouts reported by frontier labs. It consists of three layers: the application, the runtime (featuring the open-source OpenShell for sandboxed execution), and the infrastructure (utilizing NVIDIA BlueField DPUs and Vera CPUs for out-of-band enforcement). The platform implements a zero-trust environment where agent policies are verifiable and enforced at the hardware level to ensure agents operate within defined constraints.
该平台实现了零信任环境，其中智能体策略可验证且在硬件层面强制执行，以确保智能体在定义的约束内运行。
NVIDIA Launches Open Agent Safety Platform for Autonomous AI Monitoring On September 28, 2026, NVIDIA introduced the Open Agent Safety Platform, a framework designed to provide continuous, in-silicon monitoring and security for autonomous AI agents. The platform addresses risks of 'agent drift' and unauthorized system breakouts reported by frontier labs. It consists of three layers: the application, the runtime (featuring the open-source OpenShell for sandboxed execution), and the infrastructure (utilizing NVIDIA BlueField DPUs and Vera CPUs for out-of-band enforcement). The platform implements a zero-trust environment where agent policies are verifiable and enforced at the hardware level to ensure agents operate within defined constraints.