Ensure the Security of Autonomous AI Agents with NVIDIA OpenShell

Autonomous agents represent a new milestone in the development of AI. Systems can now not only generate responses but also perform actions such as reading files, using tools, and executing workflows in corporate systems.

As the capabilities of agents increase, so does the risk at the application level. To address this issue, NVIDIA is developing the OpenShell runtime environment. It is part of the NVIDIA Agent Toolkit and serves as a secure, inherently open environment for the operation of autonomous agents.

OpenShell provides isolation for each agent in its own sandbox, separating application-level operations from infrastructure-level policies. This means that security policies are inaccessible to the agent and are enforced at the system level, preventing the possibility of policy violations or data leaks.

With OpenShell, enterprises can separate agent behavior, policy definition, and enforcement, creating a unified policy for monitoring the operations of autonomous systems. Agent coding and workflow execution occur under the same policies, regardless of the operating system.

NVIDIA collaborates with security partners such as Cisco and Microsoft Security to align the management of agent execution policies in corporate environments. OpenShell also offers users the ability to customize agent behavior and data management according to their specific requirements.

NVIDIA NemoClaw provides an open reference for creating self-improving personal AI agents. This allows users to tailor security settings to their needs, providing greater control over agent behavior.

Both projects, OpenShell and NemoClaw, are in the early stages of development, and NVIDIA is actively working with the community to safely scale autonomous agents in accordance with global security standards.