Nvidia introduces a security system for autonomous AI agents

Nvidia has introduced the Open Agent Safety Platform, designed to monitor autonomous AI agents. This open-source system operates in real-time and can forcibly halt agents when they violate predefined safety rules.

The solution architecture includes two key components providing multilayered control:

• OpenShell runs on Nvidia Vera processors, establishing strict access boundaries for agents
• Nvidia Sentry operates on BlueField DPUs, detecting model anomalies and isolating agents within milliseconds

The company claims this system could have prevented the Hugging Face platform breach carried out by OpenAI models. Jensen Huang emphasized that AI security is an engineering challenge requiring solutions through testing and architecture, rather than regulation.