AI semiconductor company Nvidia has unveiled a tool designed to prevent autonomous AI agents from exhibiting unexpected behaviors that bypass human control.

On the 28th local time, Nvidia introduced the "Nvidia Open Agent Safety Platform," which provides control functions across the entire process from the testing stage of AI agents to their actual deployment.

The platform consists of "OpenShell," an open-source software, and "Sentry," a hardware monitoring device.

OpenShell sets the behavioral boundaries of AI agents on the central processing unit (CPU) and controls them in real time, while Sentry monitors operations on the data processing unit (DPU) in milliseconds (thousandths of a second) to immediately isolate and halt AI agents if they cross established boundaries.

CEO Jensen Huang pointed out, "The remarkable potential of AI can only be realized if the problem of AI safety is solved," adding, "As AI capabilities advance, safety technology must accelerate at an equal pace."

Nvidia's release of this platform is interpreted as a response to the growing number of recent cases where AI agents bypassed control and breached external systems without explicit human instructions.

More than 100 major global technology companies participated in the development of this platform, including Anthropic, Microsoft, Salesforce, Palantir, and SpaceXAI.

(Photo: AP, Yonhap News)