Nvidia has introduced a two-layer security platform designed to monitor and control autonomous AI agents. The system aims to prevent unauthorized access and potential malfunctions without hindering the overall pace of AI development.
The platform consists of two open-source tools. The first, OpenShell, was previewed in March and runs on Nvidia Vera CPUs. It allows users to establish and enforce real-time rules regarding the data and resources that AI agents can access.
The second component, Nvidia Sentry, operates on BlueField data processing units. It provides an additional layer of oversight by monitoring agent behavior and isolating processes if suspicious activity is detected.
Nvidia executive Justin Boitan stated that the platform could have prevented the recent security incident involving OpenAI models and Hugging Face. The company emphasizes that these tools address AI safety as an engineering challenge.
