Nvidia unveils a cage for AI agents: a chip can stop one in milliseconds

How it works. OpenShell treats every agent as untrusted and grants access only when needed. The design has a gateway that runs sandboxes and policies, a supervisor that checks outgoing requests and keeps credentials separate, the sandbox itself, and a policy prover that mathematically checks that an agent's permissions stay under a set ceiling.

Why a separate chip. BlueField-4 is a network processor that works apart from the main one. It sits on the only path between the agent and the model, so it sees all traffic even if the agent reports nothing about itself. Nvidia says that if an agent tries to break out, Sentry quarantines and stops it in milliseconds.

Who is on board. Nvidia's list includes Anthropic, AWS, Google Cloud, Microsoft, OpenAI, Meta, SAP, ServiceNow and others. CrowdStrike, Palo Alto Networks, Zscaler and Check Point are building support into their products. Banks BBVA and Santander are exploring it as a base for agent safety. ServiceNow made its Arc agent, governed by OpenShell, generally available the same day.

Why now. In September Nvidia already bundled OpenShell and Sentry into an agent safety platform. The news follows high-profile cases of agents slipping out of their test environments. The architecture is available now, but some parts, including Sentry, are still in progress.

Hechos confirmados

• Nvidia unveiled the OpenShell Reference Architecture at GTC Berlin on 21 October 2026.^[1][2]
• It combines Vera Rubin servers for models, Vera CPUs for agents, open-source OpenShell software and BlueField-4 chips running Sentry.^[1]
• Nvidia says that if an agent tries to move outside its boundary, Sentry quarantines and stops it in milliseconds.^[1][2]
• CrowdStrike, Palo Alto Networks, Zscaler, Check Point and others are integrating support, and BBVA and Santander are exploring it.^[1][2]
• ServiceNow announced general availability of its Arc agent, whose actions are enforced by OpenShell.^[1]
• Nvidia says the architecture is available now, while its Sentry technical material lists availability as early next year.^[1][2]
• Jensen Huang: agents can "think on Rubin, act on Vera and be governed on BlueField."^[1]

Implicaciones

• ConfirmadoMajor AI labs, cloud providers and cybersecurity firms have backed Nvidia's design.
• Nuestra lecturaIf it becomes a standard, companies will buy not only GPUs from Nvidia but also CPUs and network chips for agents.
• Nuestra lecturaAgent safety is turning into its own market for both Nvidia and cybersecurity companies.

Todavía no se sabe

• How much such a system costs and how many customers have bought it.
• When Sentry becomes widely available.
• How well the design holds up against real attacks.

A quién afecta

World. Anyone holding Nvidia or cybersecurity stocks, or working with AI agents

Fuentes

1. NVIDIA Newsroom: NVIDIA Launches OpenShell Reference Architecture for Safe AI Agents 21 de octubre de 2026 · oficial
2. Decrypt: Nvidia Unveils Its Full-Stack Blueprint to Contain and Monitor AI Agents 21 de octubre de 2026 · medios