BBVA Partners With Nvidia to Embed Security Into AI Agents at the Architecture Level
BBVA Partners With Nvidia to Embed Security Into AI Agents at the Architecture Level #
BBVA has announced a collaboration with Nvidia to examine how Nvidia’s newly introduced OpenShell Reference Architecture can be applied to AI agent safety, security, and governance. BBVA is among the first financial institutions to engage with the framework, which Nvidia unveiled at its GTC Berlin conference on October 21, 2026.
The collaboration stems from BBVA’s push to scale AI agent deployment across its organization, covering both customer-facing services and internal processes. The bank’s stated concern is maintaining oversight and control as agents become more autonomous and take on longer, multi-step tasks. BBVA said OpenShell offers a way to build security and governance into agents at the design stage rather than adding controls later.
OpenShell Reference Architecture is a full-stack hardware-and-software system that divides agent workloads across three computing layers. Vera Rubin infrastructure runs the underlying AI models; Vera CPUs manage agent harnesses and tool interactions within OpenShell’s sandboxed runtime; and BlueField-4 data processing units run Nvidia’s Sentry security software for continuous, out-of-band monitoring and in-silicon policy enforcement. Nvidia says Sentry can quarantine agents that breach defined software boundaries in milliseconds, and the system can increase agent capacity by up to 20 percent.
Nvidia CEO Jensen Huang said at the launch that safety and control must be built into the computing architecture itself as enterprises prepare to deploy billions of AI agents. The architecture is available now and is compatible with AWS, Google Cloud, and Microsoft Azure, as well as enterprise operating systems from Canonical, Red Hat, and Microsoft.
BBVA is not the only Spanish bank involved: Santander is also exploring OpenShell as a governance framework for its own agent deployments. A group of cybersecurity vendors, including CrowdStrike, Palo Alto Networks, Cisco, Fortinet, Check Point, Okta, and Zscaler, are integrating support for the architecture into their products. SAP is connecting it to its Joule Studio runtime, and model developers including Anthropic, Google, Meta, and OpenAI have confirmed their models can run on OpenShell infrastructure.
BBVA has been extending agent use into claims management, financial crime prevention, and retail banking, and has said traceability and risk management are requirements for those deployments. The Nvidia collaboration gives it a hardware-rooted framework to test whether those requirements can be met at scale.
Copy link