Nvidia unveils new security layer to strengthen safeguards for AI agents

Yesterday 17:15

By: Azzat Manal

Nvidia unveils new security layer to strengthen safeguards for AI agents

Nvidia has introduced an open-source security platform designed to give developers and businesses greater control over autonomous artificial intelligence agents as these systems become increasingly capable of performing tasks across software, networks and enterprise environments.

The initiative, known as NVIDIA OpenShell, is designed to place security controls around an AI agent rather than relying exclusively on safeguards built into the model itself. The company describes OpenShell as a secure runtime that allows autonomous agents to operate inside isolated environments with explicit policies governing what they can access and execute.

The development comes as AI agents move beyond generating text and answering questions. Modern agents can read files, use external tools, execute code, interact with applications and access business systems, creating new security challenges when they are given persistent permissions or access to sensitive resources.

OpenShell addresses these risks through sandboxing and policy-based controls covering several parts of an agent’s operating environment. According to Nvidia, the system can restrict access to files, network connections and processes while maintaining an audit trail of permitted and blocked actions.

One of the central principles behind the platform is to keep critical security controls outside the AI agent itself. This means that even if an agent behaves unexpectedly or its internal safeguards are bypassed, restrictions imposed by the surrounding runtime can continue to limit what it is able to do.

The platform can also regulate network activity and prevent agents from reaching unauthorized destinations. File-system controls are designed to restrict access to declared locations, while process-level protections can block attempts to gain additional privileges or execute potentially dangerous operations.

Nvidia has argued that this approach is particularly relevant for long-running autonomous systems. Unlike conventional chatbots, agents may operate for extended periods, use multiple tools and interact with external services, increasing the potential consequences of a compromised or misconfigured system.

OpenShell is part of Nvidia’s broader effort to build infrastructure for autonomous AI. The company has also developed NemoClaw, which combines agent deployment tools with OpenShell’s sandbox and policy controls. Recent updates have expanded lifecycle management, network policies, credential protections and monitoring capabilities.

The emergence of dedicated agent-security infrastructure reflects a broader shift in the AI industry. As companies experiment with systems capable of carrying out multi-step tasks with limited human supervision, controlling permissions, credentials and access to corporate data is becoming an important part of deployment strategies.

For businesses, the challenge is to allow AI agents enough access to perform useful work without giving them unrestricted control over sensitive systems. Nvidia’s approach seeks to address that balance by separating the agent’s capabilities from the security boundaries imposed by the environment in which it operates.

The growing focus on agent security also highlights how the risks associated with artificial intelligence are changing. As AI systems gain the ability to take actions rather than simply provide information, protecting the surrounding infrastructure can become as important as evaluating the model itself. OpenShell represents one attempt to build that additional layer of protection into the architecture used to deploy autonomous agents.