Spanish banking group BBVA (BME: BBVA) has entered a collaboration with semiconductor leader Nvidia (NASDAQ: NVDA) to examine ways of incorporating robust security and governance measures into artificial intelligence agents starting at the earliest stages of their development.
The partnership focuses on assessing the potential of Nvidia’s newly introduced OpenShell Reference Architecture, which was presented at the company’s GTC Berlin conference.
This effort aligns with BBVA’s broader push to expand the deployment of AI agents throughout its operations while preserving clear oversight and operational control.
As AI agents become more autonomous and capable of handling increasingly complex sequences of tasks, organizations face a growing need to define precise boundaries on their behavior and to retain the ability to intervene if those boundaries are breached.
Nvidia’s OpenShell Reference Architecture addresses this challenge through mechanisms that actively regulate agent actions and can step in when predefined safety thresholds are crossed.
The announcement builds on existing open-source AI security initiatives that already bring together participants from various parts of the technology ecosystem, including BBVA itself.
These capabilities hold particular importance in highly regulated industries such as financial services, where full traceability of decisions and rigorous management of risks are essential requirements.
For BBVA, working with Nvidia offers a chance to investigate fresh methods that strengthen security and governance from the initial design phase of AI agents.
The collaboration fits within the bank’s ongoing strategy to industrialize the creation and rollout of such systems across the entire organization, with the dual aims of enhancing customer experiences and accelerating internal processes.
Valentín Sánchez, BBVA’s global chief security officer, emphasized the priority placed on safe expansion: the bank is extending AI agents organization-wide, and doing so securely remains a core concern.
As these systems gain greater independence, stronger tools are required to monitor their activities and preserve control. \
Engaging with technology industry partners like Nvidia, he noted, allows BBVA to stay informed about the most advanced approaches to agent security and governance so it can continue progressing in that direction.Aser Blanco, Nvidia’s global lead for the banking sector, highlighted the practical requirements for financial institutions.
To enable banks to put their own proprietary knowledge and data to work through AI agents, clear operational limits and supervisory mechanisms must accompany every increase in system autonomy.
Through the collaboration, BBVA will explore how to enforce those limits while agents are running, supporting the bank’s goals of improving customer service and employee productivity without sacrificing control.
AI agents typically carry out multi-step tasks by chaining together different actions, ranging from retrieving information to invoking external tools.
Because of this, security cannot rely solely on safeguards established during the initial design process; it also depends on the capacity to continuously observe the agents’ decisions and actions as they execute.
OpenShell Reference Architecture meets this requirement for ongoing supervision by pairing the infrastructure that runs the agents and manages their interactions with other tools together with independent security systems.
These systems can detect and contain any activity that exceeds the established boundaries.
The underlying software is open source and is available through Nvidia’s website and on GitHub.
This joint exploration comes at a time when financial institutions are seeking practical ways to harness the productivity benefits of agentic AI while meeting stringent standards for risk management, compliance, and customer trust. By focusing on embedding controls from the design stage onward and combining them with runtime monitoring, the collaboration aims to help BBVA scale its AI capabilities responsibly.
