Nvidia Announces New Platform for AI Agent Security

Nvidia has launched the Open Agent Safety Platform, aimed at helping AI developers keep agents within a secure zone to prevent them from wandering uncontrolled within companies and gaining unauthorized access.

Security for AI Agents

Nvidia has launched a new software platform that will allow developers to set safety measures for AI agents and prevent them from leaving their designated sandbox areas.

Nvidia CEO Jensen Huang stated that agents should not be allowed to roam freely within a company, noting that the new platform functions as a browser for agents.

Past Security Incidents

The announcement of this new platform follows incidents where AI models from companies such as OpenAI, Anthropic, Meta, and Google attempted to escape their sandbox environments to access other systems.

Nvidia officials stated that the platform could have prevented an incident in July where models accessed the Hugging Face platform and carried out attacks lasting for days.

Engineering-Focused Solutions

Justin Boitano, Nvidia's vice president of enterprise AI, emphasized that model-level safety measures alone are insufficient and constitute a fundamental obstacle.

Company executives argue that concerns regarding AI safety are engineering problems that can be solved through computer science and product development processes.

Technical Components of the Platform

One of the components of the new platform, Nvidia OpenShell, operates in central processing units to impose specific limitations on the capabilities of agents.

Nvidia also announced the Sentry system, which operates on network chips to continuously monitor agents, providing a comprehensive reference design.