NVIDIA Introduces New Security System for Autonomous AI Agents
NVIDIA has launched a new artificial intelligence security platform designed to help companies control and monitor autonomous AI agents as these systems become capable of performing increasingly complex tasks.
The new NVIDIA Open Agent Safety Platform combines software and hardware-based security mechanisms designed to place enforceable boundaries around AI agents.
NVIDIA announced the platform on September 28, 2026, amid growing attention around AI agents that can write code, access tools, interact with computer systems and operate for extended periods with limited human intervention.
The company says the platform is designed to provide security controls from AI-agent testing through deployment.
What Is an AI Agent?
An AI agent is an artificial intelligence system capable of doing more than simply generating an answer.
Depending on how it is configured, an agent can receive a goal, plan a series of actions, use software tools, access files, make API calls, write code and continue working as new information becomes available.
This capability can make AI systems useful for software development, research, business automation and many other tasks.
However, giving an AI system the ability to interact with external systems also creates security challenges.
An agent that has excessive permissions could potentially access information or systems beyond what it was intended to use.
NVIDIA's New Platform Has Two Main Components
The Open Agent Safety Platform combines two major technologies: NVIDIA OpenShell and NVIDIA Sentry.
OpenShell operates as a secure runtime environment for AI agents, while Sentry provides an additional hardware-based monitoring and enforcement layer.
Together, NVIDIA says the technologies can provide controls across software, computing infrastructure and systems used to run AI agents.
OpenShell Creates a Security Boundary Around AI Agents
NVIDIA OpenShell is an open-source secure runtime designed to run autonomous AI agents inside isolated environments.
The system allows organizations to define what an agent can access, including files, networks, tools, processes and credentials.
Instead of relying only on the AI model itself to follow instructions, OpenShell places security controls outside the agent's own reasoning process.
This is important because an AI model can make mistakes, misunderstand instructions or attempt actions that were not expected by its operator.
OpenShell is designed to enforce predefined policies while the agent is running.
AI Agents Can Be Restricted Before They Start Working
According to NVIDIA, OpenShell checks an agent's permissions before it begins operating and continues enforcing those restrictions while the agent performs its tasks.
The system can control access to files, network resources, tools and other computing resources.
This creates a type of sandbox around the AI agent.
The concept is similar to security isolation used in other areas of computing, where an application is prevented from freely accessing the entire system.
NVIDIA Sentry Adds Hardware-Level Monitoring
The second major component is NVIDIA Sentry.
Sentry is designed as an independent watchdog that runs on NVIDIA BlueField-4 data processing units.
It continuously monitors agent activity and can intervene if an agent attempts to move outside its permitted boundaries.
NVIDIA says Sentry can quarantine an agent in milliseconds when its behavior violates defined policies.
The hardware-based approach provides an additional layer of protection that operates separately from the AI agent itself.
Why NVIDIA Is Building This Technology Now
The launch comes as AI agents are becoming more autonomous.
Modern agents can perform long-running tasks rather than simply responding to individual prompts.
A coding agent, for example, can inspect a software project, write files, execute commands, test code and continue modifying the project based on the results.
The more authority an agent receives, the more important it becomes to control what that system is allowed to do.
NVIDIA's platform is designed around this problem.
Recent AI Security Incidents Increased Attention
The new platform arrives after several incidents involving autonomous AI systems interacting with computer networks and online services.
Reuters reported that NVIDIA said its new tools could have prevented the high-profile breach involving AI platform Hugging Face.
The incidents have increased interest in technical systems capable of restricting AI agents before they can perform unauthorized actions.
NVIDIA is presenting its approach primarily as an engineering and security problem rather than relying only on the AI model to behave correctly.
AI Safety Is Moving Outside the Model
One of the key ideas behind NVIDIA's platform is that an AI system should not necessarily be trusted to police itself.
An AI model may be highly capable while still making errors or interpreting a task in unexpected ways.
External security controls can therefore establish boundaries that remain in place regardless of what the model decides to do.
This approach separates the AI's ability to reason from the permissions it receives.
Zero-Trust Approach for AI Agents
NVIDIA's OpenShell technology uses a security approach in which access is denied by default and permissions are granted according to predefined policies.
This can help organizations prevent an AI agent from automatically receiving broad access to sensitive files, networks or credentials.
For enterprise deployments, such controls can become especially important because AI agents may eventually interact with databases, customer information, internal applications and cloud infrastructure.
The Platform Is Designed to Work With Different AI Models
OpenShell is designed to be model-agnostic and can work with different AI models and agent frameworks.
NVIDIA says the software can also be extended to third-party computing platforms, including hardware based on Arm and Intel architectures.
This means the security concept is not limited to a single AI model or a single type of computer.
The broader objective is to create a security layer that can operate across different AI-agent environments.
Major Technology Companies Are Supporting the Initiative
NVIDIA said organizations across the AI ecosystem are participating in the initiative.
The companies and organizations listed by NVIDIA include AI companies, cybersecurity firms, infrastructure providers and enterprise software companies.
Participants include Anthropic, Cisco, CrowdStrike, Dell Technologies, HPE, Hugging Face, Microsoft, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, ServiceNow and SpaceXAI, among others.
The participation reflects the growing need for security standards as AI agents become more deeply integrated into enterprise systems.
AI Agents Could Become Part of Everyday Business
AI agents are increasingly being developed for tasks that traditionally required multiple software tools and human employees.
Companies are exploring agents for coding, customer service, research, data analysis, cybersecurity and business operations.
If these systems become widely deployed, businesses could potentially operate large fleets of agents simultaneously.
Managing those systems would require tools capable of monitoring permissions, actions and unusual behavior across thousands or even millions of AI processes.
Why Human Oversight Still Matters
Security software can restrict an AI agent's access, but it cannot automatically eliminate every possible AI failure.
An agent can still make incorrect decisions while operating inside its permitted environment.
For example, an AI system could misunderstand a legitimate task and make an inappropriate change without technically violating its access policy.
Organizations therefore need a combination of model evaluation, security controls, monitoring and human oversight.
NVIDIA's Approach Combines Software and Hardware
The Open Agent Safety Platform is notable because it does not rely exclusively on software.
OpenShell creates a controlled runtime environment, while Sentry adds an independent monitoring layer at the hardware level.
NVIDIA says this layered design can provide protection even if an agent attempts to work around application-level restrictions.
The company is effectively creating a security boundary around the AI agent rather than expecting the model to enforce every safety rule itself.
Could This Become a Standard Layer for AI Computing?
The rapid growth of autonomous AI is creating a new security category.
Traditional cybersecurity focuses on protecting computers, networks, applications and users.
AI agents introduce another type of actor: software that can make decisions, use tools and perform actions on behalf of a person or organization.
Security systems will increasingly need to determine not only whether a user or application is authorized, but also what an autonomous AI system is allowed to do at every stage of a task.
The AI Security Market Is Expanding
As businesses adopt AI agents, demand for AI security technologies is expected to grow alongside them.
Companies will need systems that can monitor agent behavior, protect credentials, isolate workloads and detect unusual activity.
They may also need detailed records showing which agent performed which action and why that action was permitted.
NVIDIA's platform is aimed at this emerging infrastructure layer.
What NVIDIA's Launch Means for the Future of AI
The launch demonstrates how the AI industry is moving from simple AI assistants toward autonomous systems that can interact with real computing environments.
As AI agents receive greater access to tools and systems, controlling their permissions becomes increasingly important.
NVIDIA's OpenShell and Sentry technologies are designed to provide those controls at different levels of the computing stack.
The company is also making OpenShell open source, allowing developers and organizations to inspect, extend and use the technology in different environments.
What Happens Next?
The effectiveness of AI-agent security systems will ultimately depend on how they perform in real-world deployments.
Organizations will need to test whether security boundaries can be maintained without preventing legitimate AI tasks from being completed.
Developers will also need to establish clear policies defining what individual agents are allowed to access and change.
As AI agents become more capable, these policies may become an essential part of enterprise computing infrastructure.
Conclusion
NVIDIA's Open Agent Safety Platform represents a major effort to build a dedicated security layer for autonomous artificial intelligence.
Its two main technologies, OpenShell and Sentry, are designed to control AI agents through software isolation and hardware-based monitoring.
The development comes at a time when AI agents are gaining the ability to perform increasingly complex tasks across computers, networks and business systems.
Rather than relying entirely on an AI model to follow its own safety instructions, NVIDIA's approach places enforceable boundaries around the system.
If autonomous AI becomes a standard part of business and computing, technologies capable of monitoring and restricting agent behavior could become an important part of the infrastructure supporting the next generation of artificial intelligence.
Journalist: Vijay Singh