Nvidia unveils OpenShell at GTC Berlin, deploying BlueField-4 to isolate rogue AI agents
The new design uses BlueField-4 units to quarantine rogue agents, with support from major cloud providers and security vendors.
Key points
- Nvidia released the OpenShell Reference Architecture at GTC Berlin, available immediately.
- BlueField-4 data processing units run Sentry software to stop agents breaching boundaries.
- Partners include Microsoft, Google, AWS, Anthropic and security firms like CrowdStrike.
Nvidia has published a complete hardware and software design for containing AI agents within strict operational limits, targeting enterprise data centres and edge computing environments. The OpenShell Reference Architecture was announced at the GTC Berlin conference and is available for immediate implementation. This release extends the vendor’s previous open-source runtime software into a full-stack infrastructure solution, as TechSec Observer reported on 2026-10-21.
Root cause
The architecture addresses the risk of AI agents operating outside defined parameters by isolating their execution environment. Nvidia’s Jensen Huang stated that safety and control must be embedded directly into the computing architecture as enterprises prepare to deploy billions of agents. The design separates the models, which run on Vera Rubin systems, from the agent-driving software, which operates on Vera CPUs, creating a clear boundary for enforcement.
Attack path
If an agent attempts to move outside its designated software boundary, the system detects the deviation using Nvidia Sentry. This software runs on BlueField-4 data processing units and monitors agent behaviour from outside the main system. According to Nvidia, Sentry quarantines and stops the offending agent in milliseconds, preventing further unauthorized actions while potentially raising overall agent capacity by up to 20%.
Affected versions
The reference architecture supports a wide range of hardware and software ecosystems. It covers edge machines including RTX Spark laptops, DGX Spark and DGX Station systems. Agents from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence are compatible. Security companies such as 1Password, Check Point, CrowdStrike, Cisco, Fortinet, Okta, Palo Alto Networks and Zscaler are adding support for the framework.
Mitigation
- Deploy agents using the OpenShell Reference Architecture to enforce hardware-level boundaries.
- Utilise Nvidia Sentry on BlueField-4 units to monitor and quarantine rogue behaviour.
- Integrate with supported security vendors like CrowdStrike or Zscaler for enhanced governance.
- Ensure agent actions are logged, as demonstrated in Anthropic’s partnership with OpenShell.
What it means for organisations
Engineering and security teams should evaluate whether current AI deployments lack sufficient hardware-enforced boundaries. Organisations must determine if their existing agent frameworks can be integrated with OpenShell or if new infrastructure is required. Risk managers need to understand how Sentry’s quarantine mechanisms align with internal incident response protocols. Procurement should verify that planned hardware purchases, such as Vera Rubin systems, support this architecture to ensure long-term safety compliance and vendor interoperability.
Background: AI agents
AI agents matter for security because they move beyond generating text to executing code and accessing systems. This autonomy means a single compromised prompt can trigger cascading failures across your infrastructure, requiring new controls for intent verification and action boundaries rather than just content filtering.
Read the full guide: Why AI Agents Matter for Security: Autonomy Risks
What to do and how to stay safe: Nvidia
- Review current AI agent deployments for potential boundary violations.
- Assess compatibility of existing infrastructure with OpenShell requirements.
- Monitor vendor updates for Sentry and BlueField-4 integration details.
- Consult with security teams on logging and governance policies for agents.
General security guidance from the TechSec Observer newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Is OpenShell available now?
Yes, Nvidia announced the OpenShell Reference Architecture at GTC Berlin and it is available immediately.
Which companies are supporting this architecture?
Anthropic, OpenAI, Meta, Microsoft, Google Cloud and AWS are working with it, alongside several security vendors.
How does the system stop rogue agents?
Nvidia Sentry runs on BlueField-4 units to quarantine and stop agents breaching boundaries in milliseconds.
