BBVA is exploring NVIDIA’s OpenShell architecture to strengthen AI agent security as the Spanish bank moves to scale autonomous systems across its operations.

As banks move beyond AI assistants towards systems that can take actions independently, security is becoming a central challenge. These agents can retrieve information, interact with software tools, and complete multi-step tasks. However, their growing autonomy also raises questions about access rights, oversight, and operational risk.

Spanish banking group BBVA is addressing this challenge through a new collaboration with NVIDIA. Announced on October 21, 2026, the initiative will explore NVIDIA’s OpenShell Reference Architecture as a way to strengthen the safety, security, and governance of AI agents across the bank. The architecture was unveiled at NVIDIA’s GTC Berlin conference. Source: BBVA Source: NVIDIA

The announcement reflects a wider shift in financial services. Banks are looking for ways to deploy AI at scale, but they must also ensure that these systems operate within defined boundaries. In banking, an agent that accesses the wrong information or performs an unauthorized action could create security, compliance, and customer-trust problems.

For BBVA, the objective is to explore ways to embed safeguards into AI agents while the bank expands their use across customer services and internal processes. The announcement describes an exploration initiative, not a confirmed bank-wide deployment of NVIDIA’s architecture.

Why BBVA Is Exploring AI Agent Security

Traditional AI assistants often respond to prompts or generate recommendations. AI agents can go further by selecting tools, accessing information, and carrying out sequences of actions to complete a task.

That ability can improve productivity. It can also create risks that are harder to manage when agents operate across multiple systems.

A model may produce an unsuitable answer, but an agent can potentially turn a flawed decision into an action. Depending on its permissions, it might access sensitive files, interact with an external service, or initiate a process that affects business operations.

Consequently, financial institutions need more than safeguards applied when a model is developed. They also need controls that govern what an agent can access and do while it is running.

BBVA says its collaboration with NVIDIA will help it explore approaches to monitoring agent activity and maintaining control as these systems become more autonomous. Valentín Sánchez, BBVA’s Global Chief Security Officer, said security is a priority as the bank expands its use of AI agents.

The bank’s focus is particularly relevant because financial services depend on traceability, risk management, and clearly defined responsibilities. Any move towards greater automation must account for those requirements.

The collaboration also fits BBVA’s broader plan to industrialize AI. Rather than building isolated experiments, the bank wants to develop and deploy agents across its organization to improve customer experience and streamline internal processes.

What Is NVIDIA OpenShell?

NVIDIA’s OpenShell Reference Architecture is designed to help organizations build, run, observe, and govern AI agents within defined security boundaries.

It combines infrastructure for running agents and managing their interactions with tools with separate security mechanisms that can detect and contain activity that exceeds established policies. NVIDIA says the approach is intended to make agent behavior easier to control during execution, rather than relying only on safeguards built into the model or application.

The underlying OpenShell secure runtime software is open source. NVIDIA’s wider reference architecture brings together computing infrastructure, agent runtime controls, and security enforcement capabilities. Source: NVIDIA Newsroom

This distinction matters because an AI agent does not necessarily follow one fixed sequence of steps. It may choose different tools or approaches as it works towards a goal. Its actions, therefore, need to remain within authorized limits even when its execution path changes.

OpenShell is designed to provide that boundary. Its controls can restrict access to resources and help govern interactions with external tools. The wider architecture also incorporates independent security mechanisms for monitoring and enforcing policies.

For banks, such an approach could support a more structured way to manage AI agents that interact with enterprise systems. However, the effectiveness of any implementation will depend on how organizations configure permissions, define policies, monitor activity, and respond to incidents.

How the Collaboration Could Matter for Banking

The collaboration centers on exploring how the architecture could support AI agents within BBVA. It does not establish that the bank has already deployed OpenShell across its systems or achieved specific security improvements.

Even so, the initiative highlights three important areas for banks adopting agentic AI.

1. Controlling access to sensitive information

Banking systems hold sensitive customer, transaction, and business information. AI agents may need access to some of these resources to perform useful tasks, but unrestricted access would create unnecessary exposure.

A policy-based approach can help define which information and tools an agent may use. For example, a system designed to assist employees with a customer query should receive only the permissions needed for that task.

The principle is straightforward: an agent should not automatically gain broad access simply because it can use a tool.

Banks would still need to connect these controls to their identity systems, data permissions, and existing security policies. They would also need to review permissions as business needs change.

2. Monitoring actions while agents operate

Security checks performed before deployment cannot anticipate every action an agent might take during a complex task.

Runtime monitoring can provide another layer of oversight. It can help identify attempts to cross defined boundaries and support intervention when activity violates policy.

This matters in financial services because an agent’s actions may affect more than the immediate user interaction. Depending on its role, it could touch internal workflows, customer information, or systems used by operational teams.

Independent controls can reduce reliance on the agent to report its own behavior accurately. However, monitoring alone does not guarantee that every threat will be detected or that every harmful action will be prevented.

3. Building governance into AI deployment

Banks must establish who can authorize an AI agent, which tasks it can perform, and when a human must intervene.

These decisions should form part of the deployment process rather than being added after an agent becomes operational. Institutions also need audit trails, incident-response procedures, and clear accountability for automated actions.

BBVA’s collaboration with NVIDIA reflects this emphasis on governance alongside technical capability. The bank is exploring how security controls can support wider AI adoption without giving up oversight.

The wider lesson is that scaling AI agents requires more than access to capable models. It also requires the infrastructure and operational processes to manage their behavior.

BBVA’s Wider AI Strategy Is Moving Towards Scale

The NVIDIA collaboration is part of a broader transformation at BBVA. The bank has been developing a common approach to building, deploying, and governing AI agents across its business.

In October 2026, BBVA outlined its “The Frame” model, a common framework intended to help teams create and manage agents more consistently. The bank said the framework aims to speed up deployment, embed governance, and preserve flexibility in its choice of technology providers. Source: BBVA

BBVA has also reported substantial employee use of generative AI. According to the bank’s published information, 75% of its employees use generative AI regularly and save an average of around 2.4 hours per week on repetitive, lower-value tasks.

The bank has described other AI applications across its operations. These include support for relationship managers, risk analysis under professional supervision, and software development. BBVA has reported productivity improvements of close to 28% in software development.

These figures are company-reported results. They indicate the scale of BBVA’s AI ambitions, but they do not measure the performance or security of the OpenShell collaboration itself.

The bank has also discussed plans to evolve its Blue financial assistant into a multimodal personal adviser. BBVA said it intends to roll out the assistant gradually to individual and business customers in 2027.

As these initiatives expand, governance becomes more important. A common framework can help an institution establish consistent development and deployment practices, while runtime controls can help enforce boundaries as agents operate.

The NVIDIA collaboration therefore complements BBVA’s broader strategy: expand AI use while developing the systems needed to manage it responsibly.

What NVIDIA Brings to the Initiative

NVIDIA is positioning OpenShell as part of a broader infrastructure approach to enterprise AI agents. The company’s reference architecture combines agent runtime software with computing and security components designed to support controlled execution at scale.

For BBVA, working with NVIDIA provides an opportunity to assess how these capabilities might fit its banking environment.

Aser Blanco, NVIDIA’s Global Banking Lead, said clear boundaries and oversight should accompany greater agent autonomy. The collaboration will explore how to enforce those boundaries while supporting the bank’s goals for customer experience and employee productivity.

The initiative also reflects a broader technology-industry effort to establish approaches for safer AI agent deployment. NVIDIA has announced work with organizations across infrastructure, software, cloud, and cybersecurity.

However, a reference architecture is not a complete governance programme by itself. Banks still need to determine which tasks are suitable for automation, how permissions should be assigned, how activity will be audited and who is responsible when systems fail.

They must also test controls against realistic scenarios, including compromised credentials, unexpected agent behaviour and attempts to access unauthorised resources.

Open-source availability may give organisations flexibility to inspect and adapt the runtime software. Yet open source does not remove the need for secure configuration, ongoing maintenance, independent testing and operational oversight.

The Risks Banks Must Still Manage

The promise of AI agents is that they can handle complex tasks with less manual intervention. Their challenge is ensuring that greater autonomy does not create unacceptable operational risk.

For banks, that challenge includes several connected concerns.

First, permission design must be precise. Excessively broad access can expose sensitive data or systems, while poorly designed restrictions can prevent agents from completing legitimate tasks.

Second, human oversight must be proportionate to risk. An agent that summarises internal information may require different controls from one authorised to trigger a financial or customer-impacting process.

Third, auditability must support investigations and accountability. Institutions need records that explain which actions occurred, what permissions applied and how exceptions were handled.

Finally, security controls require ongoing evaluation. Agents, models, tools and business processes change over time. A policy that works in one environment may not be sufficient after an agent receives new capabilities or access.

NVIDIA’s OpenShell architecture is designed to address parts of this problem through runtime boundaries and security enforcement. BBVA’s exploration will help determine how the approach could fit into its wider AI strategy. The announcement does not yet provide results from a production deployment, independent performance tests or quantified risk reductions.

What the BBVA–NVIDIA Collaboration Signals for Fintech

The significance extends beyond one bank and one technology provider. Financial institutions are increasingly looking at AI agents as a way to improve service, accelerate internal work and connect systems that previously required manual coordination.

But the path from experimentation to production depends on trust. Banks need to demonstrate that automated systems follow defined permissions, remain observable and operate under accountable governance.

This creates an opportunity for fintech infrastructure providers, cybersecurity companies and enterprise software firms. As financial institutions adopt agents, demand may grow for identity controls, permission management, secure execution environments, monitoring, audit trails and policy enforcement.

It also creates a more demanding standard for AI vendors. Strong model performance alone may not be enough. Banks will need evidence that the surrounding systems can manage access, constrain actions and support investigation when something goes wrong.

BBVA’s move to explore NVIDIA OpenShell illustrates this shift. Security is becoming part of the infrastructure required to scale agentic AI, rather than an afterthought once automation is already in place.

For financial services, the goal is not simply to make AI agents more autonomous. It is to make their autonomy usable within the controls that regulated institutions require.

What Happens Next?

The immediate next step is to assess how NVIDIA’s OpenShell Reference Architecture can support BBVA’s security and governance requirements. The announcement does not specify a deployment timetable, a confirmed production rollout or measurable outcomes from the collaboration.

The key questions will be how the bank applies policy boundaries, integrates runtime controls with existing systems and evaluates the resulting security and operational performance.

As BBVA expands AI agents across its organisation, the balance between autonomy and control will remain central. The collaboration offers a look at how a major bank is exploring that balance with an established AI infrastructure provider.

TheFinRate Takeaway: BBVA’s work with NVIDIA highlights a crucial phase in banking AI adoption. As agents move from answering questions to taking actions, banks must build security, permissions and oversight into the environments where those agents operate. The competitive advantage may belong not only to institutions that adopt AI quickly, but also to those that can deploy it safely and govern it consistently.