NVIDIA unveils OpenShell reference architecture for secure enterprise AI agents
For the head of AI platform engineering at a regional bank
NVIDIA announced the NVIDIA OpenShell Reference Architecture on Oct. 26, a design for building, containing, observing and enforcing safe AI agents on enterprise infrastructure and edge systems.1
NVIDIA's technical blog said agents can work for hours, download code, build new tools and request more privileges, so security and governance "belong in the infrastructure the agent operates in" rather than in the agent.2 In the design, NVIDIA Vera CPUs run OpenShell software, while BlueField-4 DPUs run NVIDIA Sentry in a separate hardware-isolated domain outside the agent's reach.12
According to NVIDIA, if an agent tries to move outside its software boundary, Sentry quarantines and stops it in milliseconds, and Sentry running on BlueField-4 boosts agent capacity by up to 20%.1
NVIDIA's technical blog said Sentry monitors requests, responses, networking and file activity, and verifies identity and delegated authority, without relying on agent self-reporting.2
NVIDIA said BBVA and Santander are working with it to explore the architecture as a framework for agent safety, security and governance; the source does not say whether either bank has deployed it.1
NVIDIA said ServiceNow announced general availability of ServiceNow Arc, and that ServiceNow's AI Passport carries policies to OpenShell and Sentry so enterprises can revoke an agent's access across software and hardware within seconds.1
The design puts the enforcement point outside the system the agent runs in, so an agent that changes its own software cannot disable it. Banks evaluating agent vendors can ask where each vendor's enforcement sits, and the hardware-based version described here depends on NVIDIA silicon.
Our take. The useful idea is the guard sitting on a separate chip the agent cannot touch. Demand that separation from every agent vendor, even if you never buy the Nvidia hardware sold with it.
Action. AI platform engineering: Require each agent vendor to show where its monitoring and kill switch run, and whether the agent's own runtime can disable them.
What our records say
Scale: BBVA has 1,127 AI use cases in production (as of Jul 14, 2026).1 Employee rollout: ChatGPT Enterprise began with 3,300 licences in May 2024, grew to 11,000 in May 2025, and on 12 December 2025 became a multi-year alliance extending it to all 120,000 employees in 25 countries.2 About 100,000 employees were using it in June 2026, with more than 20,000 GPTs and about 4,000 in frequent use.3
From our records
- BBVAbank · latest First European bank to join the Poppy agent protocol (Oct 2026)
- Santander USbank · latest Riley: AI lets her 'vibe code' a CRM tool herself (Oct 2026) · vendors FICO, Microsoft Copilot, ChatGPT Enterprise · enforcement 2 actions since 2016; latest CFPB consent order, $4.8 million (Dec 2020)