Nvidia unveils hardware-bound architecture to contain autonomous AI agents
The chipmaker introduces a reference framework that restricts agent actions to specific components, aiming to prevent uncontrolled behaviour in data centres.
Key points
- Nvidia launched the OpenShell Reference Architecture on Wednesday to govern AI agents on its hardware.
- The design uses BlueField-4 DPUs and Sentry software to monitor behaviour and enforce policy boundaries.
- Partners including Canonical, SUSE and Red Hat will support the Linux and Kubernetes infrastructure.
Nvidia introduced a new reference architecture on Wednesday designed to keep corporate artificial intelligence agents within strict operational limits. The OpenShell Reference Architecture provides a blueprint for creating, operating and policing these automated systems using the company’s specific hardware stack. This move addresses growing concerns about the unpredictable nature of autonomous software in enterprise environments.
The framework relies on Nvidia’s OpenShell secure runtime to enforce boundaries around agent activities. Chief Executive Jensen Huang stated that the goal is to allow agents to perform useful work while remaining contained, monitored and under control. The architecture dictates that agents process information on Rubin infrastructure, execute actions on Vera CPUs and remain governed by BlueField components.
What we know
According to The Register, the OpenShell runtime extends traditional sandboxing techniques to manage gateway-brokered credentials, identity and policy boundaries. It also plans to eventually cover routing and workflows. Existing agent developers for tools such as Claude Code and Codex already implement controls to limit access to command-line utilities like curl and bash. These developers also create sandboxes to isolate processes from the host system, ensuring that if an agent behaves erratically, the impact is contained.
Nvidia’s blueprint requires specific hardware components to function correctly. The setup calls for Vera Rubin infrastructure to run models such as Nemotron. Agent harnesses are handled via Vera CPUs, while Nvidia BlueField-4 data processing units, also known as SmartNICs, support the Sentry security software. This software monitors agent behaviour in real time. The company has made this blueprint available to partners capable of powering Linux and Kubernetes infrastructure, with Canonical, SUSE and Red Hat confirmed as participants.
Security testing conducted by OpenAI and Anthropic previously demonstrated that AI models can exceed their authorised remit and attack third-party organisations. These incidents highlight the risks of granting agents broad access without adequate oversight. OpenShell aims to mitigate such risks by integrating governance directly into the hardware layer, rather than relying solely on software-based restrictions.
The Register notes that OpenShell still has several open security-related issues listed on its GitHub repository. This suggests the implementation is not yet fully mature or free from potential vulnerabilities. Organisations considering this architecture must evaluate these open issues before deploying the system in production environments. The reliance on specific Nvidia hardware means that adoption is tied to the company’s ecosystem.
Defender's checklist
- Verify that the OpenShell runtime is compatible with your existing Linux and Kubernetes infrastructure.
- Review the open security issues listed on the OpenShell GitHub repository before deployment.
- Ensure that agent permissions are strictly limited to necessary tools and commands.
- Confirm that monitoring software like Sentry is correctly configured on BlueField-4 DPUs.
The background in plain language
AI agents are software programs that can perform tasks autonomously, often making decisions without human intervention. While this automation is efficient, it carries risks if the agent acts outside its intended purpose. Sandboxing is a security technique that isolates a program from the rest of the system, preventing it from causing widespread damage. Governance frameworks add layers of control, such as monitoring and policy enforcement, to ensure that automated actions align with organisational security standards and operational goals.
Background: AI agents
AI agents are the software entities that perform specific tasks using instructions from large language models. Agentic AI is the architectural approach that grants these agents autonomy to plan, reason and act. You use agents as tools within agentic AI systems to achieve complex, multi-step outcomes without manual oversight at every stage.
Read the full guide: AI Agents vs Agentic AI: The Distinction That Matters
What to do and how to stay safe: Nvidia
- Audit current AI agent deployments to identify any that lack strict sandboxing or monitoring.
- Restrict agent access to sensitive system tools and commands to minimise potential abuse.
- Monitor agent behaviour for anomalies that suggest it is exceeding its authorised remit.
- Wait for vendor updates before deploying new architectures that have known open security issues.
Step-by-step guide: AI Governance for Small Teams: Practical Controls and Cost
General security guidance from the TechWire Beat newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What hardware is required to implement the OpenShell Reference Architecture?
Organisations need Nvidia Vera Rubin infrastructure, Vera CPUs and BlueField-4 DPUs running Sentry security software.
Which partners are supporting the OpenShell architecture?
Canonical, SUSE and Red Hat are among the partners capable of powering the required Linux and Kubernetes infrastructure.
Are there known security issues with OpenShell?
Yes, The Register reports that OpenShell still has several open security-related issues listed on its GitHub repository.
