AI in action · Tools

Nvidia Published a Hardware Blueprint for Fencing AI Agents, With BBVA and Santander Testing It. Zscaler and Okta Rose 8% on the Day

At GTC Berlin on October 21, 2026 Nvidia released the OpenShell Reference Architecture, a design that runs an AI agent's model on one set of chips, its software and tools on another and its governance on a third, with a watchdog called Sentry that quarantines an agent within milliseconds if it tries to act outside its boundary. Agents from eight model makers, including Anthropic, Google, Meta, OpenAI and SpaceX Super Intelligence, can run on it, and eight security vendors are integrating; BBVA and Santander are exploring it for agent governance, The Next Web reported. Zscaler rose 7.84% and Okta 7.54%. It is a product, not a rule, and the banks' use of it is a test.

Oct 22, 2026

The question regulators have been asking banks all month, how an autonomous AI agent would be stopped, got an engineering answer from a chipmaker on Wednesday. At its GTC Berlin conference on October 21, 2026 Nvidia released the OpenShell Reference Architecture, which builds on OpenShell, the open-source runtime it published in September under an Apache licence. The design separates an agent into three layers on three kinds of hardware: the model runs on the company's Vera Rubin systems, the agent's software and tools on Vera CPUs, and the governance layer on BlueField-4 data-processing units, outside the system the agent can reach. Jensen Huang's formulation, as reported by The Next Web, was that agents "think on Rubin, act on Vera and are governed on BlueField," and that "AI capability and safety must advance together."

The governance component is the part of interest to anyone who runs money. Nvidia Sentry, which runs on the BlueField chips, monitors an agent's behaviour from outside the main system, inspects its requests and responses, verifies its identity, enforces access policies for data, tools and services, and, if the agent attempts an action beyond its boundary, quarantines and stops it within milliseconds, by the company's account. The company says agents from eight model makers can run on the architecture: Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence. Security vendors integrating include 1Password, Check Point, Cisco, CrowdStrike, Fortinet, Okta, Palo Alto Networks and Zscaler, with the caveat, noted in the coverage, that an announced integration is not a finished one. The September launch of the underlying platform had named Citi and JPMorgan Chase among the firms working with it; the October report adds BBVA and Santander as banks exploring OpenShell to strengthen agent safety and governance.

The market treated it as a security-software event. Zscaler rose 7.84% to $231.33 and Okta 7.54% to $208.64 on October 21, two of the largest gains in the S&P 1500, and CrowdStrike, which had expanded its own Nvidia agreement earlier in the month, was among the gainers; a severe PHP vulnerability disclosed the same day added to the bid for security names. The logic is that a reference architecture from the dominant AI hardware vendor makes agent governance a budget line, and the companies that sell identity, access and traffic inspection are the ones that fill it. None of the banks named has said what its agents do, and a bank exploring a reference design is a long way from a bank deploying agents that move money.

The reason it belongs in a column about regulation is that it is the first concrete design for the kill switch the supervisors keep asking about. The Fed, OCC and FDIC's questions to banks this month included how an autonomous system would be shut down; the ECB wants eurozone banks' action plans against AI-enabled attacks by October 31 and expects them to be able to switch systems off and restart them; Hong Kong's monetary authority said on October 16 that banks should have response measures "possibly including kill-switches." A hardware watchdog that stops an agent in milliseconds is one way to answer, and it is being sold by the company whose chips the agents run on. Whether supervisors accept a vendor's boundary as a bank's control is the next question, and it will be asked of BBVA and Santander before anyone else.

Sources: The Next Web — Nvidia turns OpenShell into a full hardware stack for safe AI agents (Oct 21, 2026); StockStory via TradingView — Zscaler and Tenable shares are soaring, what you need to know (Oct 21, 2026); FourWeekMBA — Nvidia Says Agents From 8 Model Makers Can Run on OpenShell (Oct 2026)