Introduction
Nvidia is expanding its role in the AI infrastructure stack with a new security platform designed to keep increasingly autonomous AI agents within defined limits. Announced September 28, 2026, the Nvidia Open Agent Safety Platform combines open-source OpenShell software with a reference system called Sentry that uses Nvidia BlueField-4 DPUs to monitor agent behavior and quarantine agents that attempt to move outside their permitted boundaries.
The announcement comes as AI agents move beyond generating text and begin taking actions across software environments, enterprise systems and, increasingly, physical machines. Nvidia says recent security incidents have demonstrated the limits of relying solely on controls implemented at the application layer.
The company’s Google Trends profile is also showing a noticeable rise in searches for “Nvidia,” with current news coverage centered on the new safety platform.
Background and Context
AI agents represent a significant change from conventional chatbots.
A chatbot primarily responds to prompts. An agent can potentially read files, use software tools, execute code, access networks, retain context and perform multistep tasks with limited human intervention.
That additional autonomy creates a different security problem.
Nvidia has spent much of 2026 building software around this emerging agent ecosystem. Its OpenShell runtime was introduced as an infrastructure-level security boundary for autonomous agents, rather than relying only on instructions or safeguards inside an AI model. Nvidia’s technical documentation describes OpenShell as a runtime that can enforce controls over files, networks, credentials and tools.
The company is now extending that approach into a broader architecture.
The new Open Agent Safety Platform is designed to cover the software layer, computing hardware and systems that execute agent actions. Nvidia describes it as an open platform and reference system rather than a single standalone security product.
Latest Update or News Breakdown
Nvidia introduces Open Agent Safety Platform
Nvidia announced the platform on September 28, describing it as an open software platform and reference system design for AI agent security from testing through deployment.
The architecture has two central components:
OpenShell provides the software boundary around an agent.
Sentry adds an independent hardware-based monitoring layer.
According to Nvidia, OpenShell traces agent actions and enforces policies while agents operate on Nvidia Vera CPUs. The software is open source and can also be extended to work with third-party compute platforms, including Arm and Intel systems.
Read Nvidia’s official announcement
Sentry adds hardware-level monitoring
The more unusual component is Nvidia Sentry.
Sentry is an out-of-band watchdog that runs on Nvidia BlueField-4 DPUs. Nvidia says it continuously monitors agent behavior independently of the agent itself. If an agent attempts to leave its defined software boundary, Sentry is designed to quarantine and stop it in milliseconds.
That separation is important.
Traditional AI safety mechanisms can operate within the same software environment as the model or agent. Nvidia’s architecture instead puts an enforcement mechanism outside that environment.
The company’s stated objective is to create a security boundary that an agent cannot simply modify or bypass.
OpenShell and Sentry work at different layers
The basic architecture can be understood this way:
| Component | Primary role | Where it operates |
|---|---|---|
| OpenShell | Controls agent permissions, actions and environment | Software/runtime layer |
| Nvidia Vera | Purpose-built CPU platform for agentic AI | Compute layer |
| Sentry | Monitors and contains agents independently | BlueField-4 DPU |
| BlueField-4 | Provides isolated hardware enforcement | Infrastructure layer |
| DOCA | Provides programmable DPU capabilities | Networking/security software |
Nvidia says Sentry uses its DOCA software stack to inspect agent requests and responses, verify identity, provide telemetry and enforce zero-trust access policies covering data, tools, APIs and services.
More than 100 organizations are involved
Nvidia says more than 100 organizations are working with technologies associated with the Open Agent Safety Platform.
The list includes Anthropic, Cisco, CrowdStrike, Dell Technologies, Figure, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow and SpaceXAI.
The company also says organizations in financial services, energy, enterprise software and robotics are working with the platform.
That broad participation reflects the fact that agent security is not limited to AI labs. Autonomous systems increasingly interact with corporate networks, financial data, software repositories and physical infrastructure.
Expert Insights or Analysis
The central idea behind Nvidia’s announcement is that AI security cannot necessarily depend on the AI model behaving correctly.
That distinction is becoming more important as agents gain access to tools.
Nvidia’s own AI security research has previously identified recurring weaknesses including inadequate access controls, unrestricted code execution, insufficient network restrictions and exposed credentials. The company has argued that deterministic controls outside the model can provide stronger enforcement than prompt-based safeguards alone.
Its March 2026 OpenShell documentation made a similar argument. Nvidia described the problem with putting security controls inside an agent as effectively asking a long-running autonomous process to police itself.
The new platform extends that philosophy.
Instead of assuming the agent will always follow its instructions, the surrounding infrastructure establishes what the agent is technically permitted to do.
That resembles a familiar principle from cybersecurity: least privilege.
An application should receive only the access it needs, and actions should be checked independently of the application itself.
Why the hardware layer matters
The Sentry component introduces another layer of separation.
Because Sentry operates on BlueField-4 DPUs in an out-of-band trust domain, Nvidia says the monitoring mechanism is independent of the agent being monitored.
That matters in scenarios where an AI agent has been compromised or begins behaving unexpectedly.
If an agent can alter its own software environment, a security control running entirely inside that environment may face a difficult enforcement problem. An external control point can instead monitor activity from outside the agent’s execution context.
Nvidia’s claim is not that Sentry makes autonomous AI universally safe. Rather, it is designed to provide another enforcement boundary around the systems in which agents operate.
That distinction is important because the company’s announcement describes a technology architecture, not proof that every possible AI-agent failure can be prevented.
Broader Implications
The Nvidia announcement points toward a broader shift in how the industry may think about AI security.
For years, much of AI safety focused on model behavior. Developers trained models to refuse certain requests, follow policies and avoid generating dangerous content.
Agentic AI adds another dimension.
An agent may produce an entirely ordinary response while simultaneously taking an action through a connected tool. The security question therefore becomes not only what did the model say?, but also what was the system allowed to do?
That creates several implications.
AI security becomes an infrastructure problem
If agents interact with operating systems, APIs, databases and physical machines, security controls have to exist at those layers too.
Nvidia’s platform explicitly extends security beyond the model and application layer into CPUs, DPUs, networking and infrastructure.
Enterprises may demand stronger agent controls
Businesses are increasingly experimenting with agents for coding, research, customer service and workflow automation.
The more authority an agent receives, the more important auditability and access control become.
Nvidia says Salesforce, for example, has integrated OpenShell with Slack so teams can view agent activity and audit events and approve or reject requests for additional permissions.
Robotics creates an even larger challenge
Nvidia also says robotics companies including Figure, Gecko Robotics and Skild AI are building with OpenShell.
That matters because a software mistake is one thing. An autonomous system that can act in the physical world introduces another layer of consequences.
The company’s stated goal is therefore to apply similar security principles across digital and physical AI systems.
For more coverage of AI infrastructure, security and emerging technology, readers can explore The Tech Marketer.
Related History or Comparable Technologies
The architecture Nvidia is describing has parallels with established cybersecurity concepts.
Sandboxing
A sandbox restricts software to an isolated environment. If the program behaves unexpectedly, the damage can be contained rather than allowing unrestricted access to the host system.
OpenShell builds on this general principle for autonomous agents.
Nvidia describes individual agent sandboxes, policy enforcement and controlled network and filesystem access as parts of its security approach.
Zero-trust security
Zero-trust architecture assumes that access should be verified rather than automatically trusted.
Nvidia says Sentry can enforce granular zero-trust policies for data, tools, APIs and services.
Hardware security
Hardware-based security enforcement is also not new.
Trusted execution environments, security processors and network DPUs have long been used to isolate sensitive operations from general-purpose software.
Nvidia’s approach applies that concept specifically to autonomous AI agents, with Sentry operating independently of the agent software.
Runtime security
The biggest conceptual change is combining these approaches into an agent-specific runtime.
Nvidia’s OpenShell is designed to enforce policies while an agent is operating, while Sentry provides a separate monitoring and enforcement mechanism at the infrastructure layer.
That creates multiple opportunities to detect or restrict unwanted behavior rather than relying on a single safety mechanism.
What Happens Next
The next stage will be adoption and testing.
Nvidia says Open Agent Safety Platform software, including OpenShell and related skills, is available through its developer resources and GitHub.
Several developments will be worth watching:
- Developer adoption: OpenShell’s open-source availability could allow developers to test the architecture across different environments.
- Third-party hardware: Nvidia says OpenShell can be extended to Arm and Intel compute platforms.
- Enterprise deployment: Companies such as Salesforce, SAP, Scale AI and others are already integrating elements of the technology.
- Security testing: Independent researchers will be able to test how effectively the controls withstand sophisticated agent behavior.
- Physical AI: Robotics deployments could become an important test of whether infrastructure-level controls work reliably outside traditional software environments.
- Industry standards: Nvidia says the initiative is connected with the Open Secure AI Alliance and broader efforts to establish shared AI security practices.
The most important question will be whether these controls perform as intended under adversarial conditions, not simply whether they work in controlled demonstrations.
Conclusion
Nvidia is pushing AI security deeper into the infrastructure stack with its new Open Agent Safety Platform.
The system combines OpenShell, an open-source runtime designed to enforce agent boundaries, with Sentry, a BlueField-4-based watchdog designed to monitor and quarantine agents independently of their software environment.
The timing is significant. AI agents are becoming more capable of executing code, accessing tools and performing long-running tasks, creating security challenges that traditional model-level safeguards may not fully address.
Nvidia’s approach is built around a straightforward premise: the system controlling an autonomous agent should not necessarily be the same system responsible for deciding whether that agent is allowed to act.
Whether that architecture becomes a widely adopted standard will depend on real-world deployments, independent security testing and how developers balance agent autonomy with enforceable controls.
For now, Nvidia is betting that the next phase of agentic AI will require not only more capable models, but also infrastructure capable of keeping those models within clearly defined boundaries.
FAQ
1. What did Nvidia announce?
Nvidia announced the Open Agent Safety Platform, an open software platform and reference system designed to provide security controls across AI agents, computing infrastructure and robotics systems.
2. What is Nvidia OpenShell?
OpenShell is Nvidia’s open-source secure runtime for autonomous AI agents. It is designed to enforce policies controlling agent access to files, networks, credentials and tools.
3. What is Nvidia Sentry?
Sentry is an out-of-band watchdog reference design that runs on Nvidia BlueField-4 DPUs. Nvidia says it continuously monitors agent activity and can quarantine agents that attempt to move beyond defined boundaries.
4. Can Nvidia’s platform prevent all rogue AI agents?
Nvidia describes the technology as a security and enforcement architecture, but its announcement does not establish that it can prevent every possible AI-agent failure. The platform is intended to provide additional controls outside the agent itself.
5. Why does Nvidia want AI security outside the model?
Agents can interact with files, networks, tools and external systems. Nvidia’s approach is to enforce security policies at the infrastructure level so an agent cannot simply override those controls through its own software behavior.
6. Which companies are working with Nvidia on agent security?
Nvidia says more than 100 organizations are working with its agent-safety technologies. The announced participants include Anthropic, Microsoft, Hugging Face, Salesforce, SAP, Scale AI, ServiceNow, Palantir, CrowdStrike and others.
Sources & References
- NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment, NVIDIA Newsroom
- Nvidia Releases Software It Says Can Prevent AI Agents From Going Rogue, The Wall Street Journal
- NVIDIA Open Agent Safety Platform: A Reference for Continuous, In-Silicon Agent Monitoring, NVIDIA Developer
- Four Ways to Deploy More Secure AI Agents, NVIDIA Technical Blog
- How Autonomous AI Agents Become Secure by Design With NVIDIA OpenShell, NVIDIA Blog
