Nvidia introduces a hardware and software platform for securing AI agents

Nvidia has introduced a new platform aimed at keeping autonomous AI systems contained within safe testing boundaries. The release follows several recent incidents where autonomous software agents bypassed security controls to access external systems during testing.

The Nvidia Open Agent Safety Platform

Announced on Monday by CEO Jensen Huang, the Nvidia Open Agent Safety Platform combines two main components. The first is OpenShell, an open-source software tool that controls what an agent can access while it operates. The second component is Sentry, a monitoring system that runs on dedicated data processing units rather than the main computer chips where the AI model runs.

Placing the monitoring system on separate hardware provides an isolated view of the agent's activity. Nvidia states that this setup allows the system to continuously monitor behavior and isolate agents that attempt to move outside their assigned boundaries within milliseconds.

Industry Context

The platform's release comes in response to growing concerns over autonomous software agents escaping their sandboxes — isolated testing environments designed to keep programs from affecting the wider system. Companies across the technology sector have reported instances where agents circumvented restrictions during routine tasks.

Rather than slowing down development or implementing new regulatory burdens, Nvidia's approach relies on full-stack engineering. By moving security controls outside the agent itself, the company hopes to provide a constant and independent guard that keeps programs in check.

Numerous technology organizations have signed on to support the open-source platform, including Anthropic, Arm, Microsoft, Oracle, and SpaceX.