Nvidia Launches Open Agent Safety Platform to Control AI Agents at Every Stage

Nvidia has unveiled the open Open Agent Safety Platform to prevent uncontrolled behavior by AI agents at every stage, from testing to deployment. The new system was launched after OpenAI, Anthropic, Meta and Google reported cases in which their AI models independently left the internal environment known as the “sandbox.”

Nvidia CEO Jensen Huang (Jensen Huang) called the new technology a “browser for agents” that provides an isolation system allowing access only to what an agent needs to do its job. “Agents cannot be allowed to wander aimlessly around organizations, so it is necessary to find a way to limit their activities,” Huang said in an interview with CNBC.

The platform is built around Nvidia OpenShell, a runtime environment that provides management tools to control what an agent can see, do and interact with. The controls remain in effect even when agents behave unpredictably, limiting the spread of errors. An additional hardware component, Nvidia Sentry, monitors AI behavior through networking chips without placing a load on the primary graphics or central processors. This separation of functions provides more reliable real-time control over the actions of autonomous systems.

The new technology is optimized to run on systems with Nvidia Vera processors and Nvidia BlueField DPUs, but it is also compatible with other hardware systems.