Nvidia Touts Open Agent Safety Platform with Real-Time Security Controls

Nvidia has introduced its Open Agent Safety Platform to address concerns about AI safety. The platform includes an out-of-band watchdog called Sentry, which runs on a separate chip from the agent it is watching. This design decision is meant to prevent agents from moving outside their software boundaries and circumventing security controls.

The point of having a second chip is that the agent cannot reach it, allowing for more robust security. Nvidia claims that Sentry can quarantine and stop an agent in milliseconds if it attempts to move outside its boundary.

OpenShell, the open-source software part of the platform, provides a secure runtime boundary that traces all actions and enforces policy as agents run on NVIDIA Vera CPUs. This allows for third-party compute platforms, including those from Arm and Intel, to be extended with the platform's capabilities.

Nvidia has partnered with several companies, including Anthropic, SpaceXAI, and Scale AI, which are already using the platform in various applications. For example, Anthropic uses Claude Managed Agents, which establish a security boundary by running the agent loop in a separate server from the sandboxes where their work executes.