NVIDIA launches OpenShell: hardware to stop AI agents that step out of bounds

NVIDIA is launching the OpenShell Reference Architecture, a reference design for running AI agents within fixed boundaries on a company's own infrastructure. According to NVIDIA, a dedicated security chip will quarantine and stop an agent within milliseconds if it tries to move outside the limits it has been given. The architecture is available now, but pricing has not been disclosed.

CEO Jensen Huang presented the news during his keynote at GTC Berlin on Wednesday 21 October 2026. NVIDIA Newsroom posted it on X at 12:00 Norwegian time, the same day the company published a press release. Among the companies NVIDIA says are working with the architecture are Anthropic, OpenAI, Microsoft, SAP, ServiceNow and the banks BBVA and Santander.

What NVIDIA is launching

The OpenShell Reference Architecture builds on OpenShell, NVIDIA's open-source software that sets a secure runtime boundary around autonomous agents. The new architecture extends this into a complete hardware and software platform.

NVIDIA splits the work into three layers. Vera Rubin systems run the AI models themselves. Vera CPUs run the agent harness and tool calls. BlueField-4, a dedicated networking and security chip, runs the NVIDIA Sentry security software. Sentry monitors agent behaviour continuously and outside the agent's own environment.

According to the press release, the control works like this: if an agent attempts to move outside its software boundary, Sentry quarantines and stops it within milliseconds. NVIDIA describes this as in-silicon enforcement, meaning in hardware rather than only in software the agent itself might influence. The company also says Sentry on BlueField-4 boosts agent capacity by up to 20 percent.

The architecture is designed to run on Linux and Kubernetes from Canonical, SUSE and Red Hat, on Windows via Microsoft Execution Containers, and in the clouds of AWS, Google Cloud and Microsoft Azure. It can also be used on smaller machines at the edge, such as RTX Spark, DGX Spark and DGX Station.

Who is involved

NVIDIA says enterprises can run agents from Anthropic, Cohere, Google, Meta, OpenAI and Perplexity, among others, on OpenShell infrastructure. Security companies including 1Password, Check Point, CrowdStrike, Cisco, Fortinet, Microsoft, Okta, Palo Alto Networks and Zscaler are integrating support into their agent security products, according to NVIDIA.

SAP is connecting the architecture to Joule Studio, part of SAP Business AI Platform. ServiceNow announced the same day that its ServiceNow Arc agent is generally available, with actions governed by the company's control tower and enforced by OpenShell. According to NVIDIA, organisations using ServiceNow's AI Passport will be able to revoke an agent's access across software and hardware within seconds.

The European banks BBVA and Santander are exploring the architecture as a framework for agent safety and governance. That does not mean they have deployed it. System makers Dell, Cisco, HPE, Lenovo and Supermicro are building infrastructure to the design.

"As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture," Huang says in the press release.

What has not been disclosed

NVIDIA does not disclose pricing, licensing terms or which parts of the solution require new hardware. The press release also does not explain how an agent's boundaries are defined in practice, who owns the rules, or how quarantine and stop events are logged and audited. The performance and capacity figures are NVIDIA's own and have not been independently tested. The press release also carries the usual caveat that several features will be offered "when-and-if-available".

What Norwegian leaders should note

• Control over AI agents is moving from policy documents into technical infrastructure. Organisations planning agents with access to systems and data should require that boundaries are enforced outside the agent itself, and that stops can be documented.
• Many vendors Norwegian organisations already use, such as Microsoft, SAP, ServiceNow, Cisco and Fortinet, are on the partner list. Ask them specifically what the support means for their products and when it will be available in Norway.
• The strongest control in the architecture sits in NVIDIA hardware. That may improve security, but also tighten dependence on a single vendor. Weigh this in procurement and risk assessments before agent projects are scaled up.

Sources and media

• NVIDIA Newsroom on X: OpenShell Reference Architecture announced at GTC Berlin, post published 21.10.2026 at 12:00 Norwegian time
• NVIDIA: NVIDIA Launches OpenShell Reference Architecture for Safe AI Agents, press release dated 21.10.2026
• Sources were checked on 21.10.2026.
• Illustration: AI-generated with Grok Bot / hogby.ai