英伟达发布开源AI代理安全平台，含OpenShell沙箱与Sentry硬件隔离方案
英伟达于2026年9月28日发布NVIDIA Open Agent Safety Platform，包含软件层OpenShell（Apache 2.0许可的运行时，支持代理沙箱与操作限制）和硬件层NVIDIA Sentry（基于BlueField-4 DPU的参考设计，可在毫秒级隔离越界代理）。该平台将监控置于代理无法访问的芯片上，通过Vera Rubin POD架构实现“模型前拦截”，确保代理在调用模型前即被审查与阻断。
Nvidia dropped an open safety platform built to catch and quarantine rogue agents in real time before they escape.
NVIDIA's new Open Agent Safety Platform moves AI agent monitoring onto a chip the agent can't reach.
Its software half, OpenShell, is an Apache 2.0 runtime that sandboxes agents and enforces operator limits on files, networks, tools and credentials.
Its hardware half, NVIDIA Sentry, is a reference design running on BlueField-4 data processing units that can quarantine straying agents within milliseconds.
In a Vera Rubin POD, each tray's BlueField-4 sits on the node's only route to the model, isolated from the host. Because agents must consult the model before acting, that position gives Sentry a view of every step and a kill switch.
来源：X：Rohan Paul (@rohanpaul_ai) · x.com