Nvidia Unveils Guardrails for Rogue AI Agents

Nvidia has introduced a new platform to prevent rogue AI agents from causing harm. The system consists of two parts: OpenShell, open-source software that limits what agents can do, and Sentry, an additional layer of security running at the chip level.

Nvidia's vice president of Enterprise AI, Justin Boitano, said the industry has been focusing on training good behavior into models, but this approach has limitations. The new system is a deterministic one that mediates and enforces how agents behave.

OpenShell sits between an agent and the enterprise system it can affect, allowing companies to set rules around what an agent can access and do. It verifies these boundaries before the agent executes, preventing unintended actions.