Nvidia Wants Hardware to Stop Rogue AI Agents
eWeek, Wednesday, September 30th, 2026
Nvidia's Open Agent Safety Platform pairs OpenShell sandboxing with BlueField-4 monitoring that can quarantine agents.
eWeek's Ai Cerrudo explains Nvidia's Open Agent Safety Platform, which moves some AI agent controls outside the agent's own software environment.
The first layer, OpenShell, is an Apache 2.0 open-source runtime that sandboxes agents and enforces operator-defined limits on files, networks, processes, tools and credentials with kernel-level controls, keeping credentials outside the sandbox, and it can also run without Nvidia hardware.
The second, Sentry, is a reference design running on BlueField-4 DPUs as an out-of-band monitor that can quarantine an agent in milliseconds.