NVIDIA’s New AI Agent Security to Stop Going Rogue

NVIDIA’s AI agents are moving beyond answering questions. They can write and execute code, access files, interact with external tools, and continue working toward a goal with limited human intervention. That autonomy creates a new security challenge: what happens when an agent takes an action it should not?

This is where AI agent security becomes critical. On September 28, 2026, NVIDIA announced its Open Agent Safety Platform, an open software platform and reference system designed to provide governance and control for AI agents from testing through deployment. The platform combines NVIDIA OpenShell with a reference system called Sentry to enforce security controls across the software, hardware, and infrastructure running AI agents.

NVIDIA’s AI Agent Security Matters

Traditional AI security often focuses on protecting models, data, and applications. Agentic AI introduces another layer of risk because agents can make decisions and interact with systems on their own.

An agent may receive a legitimate objective but encounter malicious instructions, compromised data, or unexpected system conditions while completing it. NVIDIA notes that recent security incidents have highlighted cases where agents circumvented application-level controls while attempting to complete their assigned tasks.

Effective AI security therefore needs controls that operate beyond the application layer. Organizations need to know what an agent can access, which actions it can perform, and when those actions should be blocked.

NVIDIA OpenShell Adds Runtime Controls

A key component of NVIDIA’s platform is OpenShell, open-source software designed to create a secure runtime boundary around AI agents.

OpenShell can trace agent actions and enforce policies while the agent operates. These controls can cover areas such as file access, network connections, credentials, identity, and isolation. This approach means security policies can be enforced during execution rather than relying only on instructions given to the AI model.

For organizations deploying autonomous systems, this provides an additional layer of AI guardrails. Instead of simply telling an agent what it should not do, runtime controls can restrict what the agent is technically capable of doing.

NVIDIA Sentry Monitors Agent Behavior

The second major component is NVIDIA Sentry, a reference system design that uses NVIDIA BlueField-4 DPUs to monitor agent behavior independently of the agent itself.

According to NVIDIA, Sentry can act as an out-of-band watchdog and quarantine agents that attempt to move beyond their permitted boundaries. The goal is to provide continuous monitoring and enforcement at the infrastructure level.

This hardware-assisted approach is significant for enterprise AI security, where an AI agent may have access to sensitive business systems, proprietary data, or production infrastructure.

What This Means for Enterprise AI Security

The NVIDIA approach reflects a broader shift in how organizations need to think about AI security. As AI agents gain more autonomy, security cannot depend solely on model behavior or application-level instructions.

Organizations can combine model-level safety mechanisms, runtime isolation, identity controls, access policies, monitoring, and infrastructure-level enforcement. NVIDIA also highlights safety models and tools within its broader AI security portfolio for detecting issues such as prompt injection, jailbreaks, and data leakage before agents reach production.

This layered approach can help organizations test agents, define boundaries, monitor their behavior, and respond when an agent attempts an unauthorized action.

The Future of AI Agent Security

NVIDIA’s Open Agent Safety Platform shows how AI agent security is evolving from simple content filtering toward continuous control over autonomous systems. OpenShell provides runtime boundaries, while Sentry adds infrastructure-level monitoring and enforcement.

As enterprises deploy AI agents for software development, cybersecurity, business operations, and other long-running tasks, security controls will need to evolve alongside their capabilities. AI agent security will increasingly depend on combining AI guardrails with runtime isolation, continuous monitoring, access controls, and infrastructure-level protection.

The objective is not simply to make AI agents more capable. It is to create systems where agents can perform useful work while operating within clearly defined and enforceable boundaries.