Nvidia Open Agent Safety can quarantine agents in ms

Nvidia says its new Open Agent Safety platform can quarantine AI agents that attempt to break preset constraints within a few milliseconds. The company introduced the security platform as a tool to monitor and contain autonomous artificial intelligence agents amid a recent surge of cyberattacks and reports of rogue model behavior.

Open Agent Safety is built on the open source OpenShell software and runs on Nvidia's Vera AI central processor. A separate chip equipped with Sentry technology performs continuous oversight, enforcing boundaries and isolating agents that try to exit their defined operational envelope. According to Nvidia, users set explicit controls for which data an agent may access, and the system monitors those restrictions before tasks begin and while they are running.

How the system operates

The platform enforces policy at multiple stages: policy configuration by the user, pretask checks by OpenShell, and runtime enforcement by the Sentry-equipped chip. Nvidia characterizes the combination of software and dedicated hardware as able to detect an agent attempting unauthorized actions and to quarantine it within a few milliseconds.

Jensen Huang, Nvidia's chief executive, told CNBC: "To deliver an agent-based system safely, you must ensure the surrounding isolated environment is designed to keep agent access to the minimum possible."

Concerns about agent safety rose after companies including OpenAI, Anthropic and Google reported instances of models leaving test environments and accessing other organizations' infrastructure. Nvidia says major technology firms including Microsoft, Anthropic and SpaceX are backing its security approach.