AI agents represent a fundamental shift in enterprise computing. Traditional generative AI systems answer questions, but agents take action. They can access data, invoke tools, generate code, interact with other agents, and make decisions on behalf of users and organizations. That capability is what makes agentic AI so powerful. It's also what makes it difficult to trust.
In conversations with customers, we've found that trust has emerged as one of the biggest barriers to deploying agentic AI in production. That makes sense. Most of us have seen generative AI produce an incorrect answer. With agents, the concern is different: they're no longer just generating the wrong answer. They can do the wrong thing. They can modify data, invoke tools, create code, and take actions at machine speed. That changes the risk equation for every enterprise considering agentic AI.
This challenge is driving an important industry conversation around how to safely govern autonomous systems. That's why NVIDIA's announcement of the NVIDIA Open Agent Safety Platform is so timely. The platform reflects an important principle: security cannot depend solely on how an agent is instructed to behave. Control must be enforced by the infrastructure surrounding the agent, a principle with significant implications for storage teams.
Data Is the New Attack Surface
As enterprises adopt agentic AI, data becomes both more valuable and more vulnerable. Agents don't simply consume information; they interact with it. They read files, modify records, generate outputs, move information between environments, and increasingly create code dynamically to accomplish tasks. Unlike traditional software, their actions are not always confined to predefined workflows.
At the same time, enterprise AI environments continue to become more distributed. Data moves between on-premises infrastructure, cloud services, vector databases, AI models, and agent frameworks. Every connection and data transfer create another opportunity for mistakes, misuse, or malicious activity.
This is why I increasingly view enterprise data as the primary security boundary for agentic AI. If organizations cannot fully trust every decision an autonomous agent might make, they must ensure the infrastructure controlling access to data remains trustworthy.
Storage infrastructure is no longer just where enterprise data resides. It is becoming an active participant in governance, cyber resilience, observability, and AI operations. As agents become more autonomous, the storage layer plays a critical role in ensuring data remains protected, recoverable, and auditable.
Infrastructure Creates Trust
The NVIDIA Open Agent Safety Platform addresses this challenge through infrastructure-enforced controls. Together, NVIDIA OpenShell and NVIDIA Sentry running on NVIDIA BlueField create multiple layers of governance, observability, and protection around autonomous agents:
▪ NVIDIA OpenShell provides a secure runtime for autonomous agents, creating isolated execution environments and enforcing policy boundaries around what agents can access and do.
▪ NVIDIA Sentry running on NVIDIA BlueField provides a hardware-isolated security domain that operates independently from the host, enabling monitoring, policy enforcement, and threat detection even if workloads become compromised.
The underlying idea is simple: an agent may propose an action, but infrastructure determines whether that action is allowed. That same philosophy applies to storage infrastructure. As agents gain more autonomy, organizations need the ability to monitor data access, enforce policy, maintain audit trails, and preserve recoverable copies of critical information regardless of what an agent attempts to do.
Data Resilience Is the Missing Layer
Security is essential, but security alone is not enough. No system is perfect – agents may make mistakes, credentials may be compromised, new attack paths may emerge. The question is not simply how to prevent every incident, but how to limit the impact and recover quickly when something goes wrong. That's where data resilience becomes critical.
First, organizations need visibility. If an agent accesses sensitive information, administrators need a complete audit trail showing what data was touched, when it was accessed, and what actions were taken. Auditability becomes a foundation for understanding, governing, and improving agent behavior.
Second, organizations need protection. Modern storage infrastructure plays an increasingly active role in cyber resilience strategies, helping detect threats, limit access to authorized users, preserve protected copies of data, and provide additional layers of defense around critical information assets.
Third, organizations need recovery. One of the lessons learned from ransomware is that prevention alone is never sufficient. Recovery must be part of the architecture. Technologies such as immutable snapshots provide trusted recovery points that cannot be altered or deleted, enabling organizations to rapidly restore data if an agent, user, or attacker causes damage.
This “detect, protect, and recover” model has become a cornerstone of cyber resilience. In an agentic world, it becomes even more important because the systems interacting with enterprise data are increasingly autonomous, dynamic, and operating at machine speed. When an agent can touch thousands of files faster than a human can react, resilient storage infrastructure becomes just as important as prevention.
IBM and NVIDIA: Building Confidence in Agentic AI
NVIDIA Open Agent Safety Platform helps establish trusted execution environments for autonomous agents through runtime isolation, policy enforcement, observability, and infrastructure-level security controls.
IBM brings complementary capabilities to the data layer. IBM Storage solutions help organizations strengthen cyber resilience through threat detection, protected copies, immutable snapshots, auditability, and rapid recovery.
Together, these capabilities help organizations move beyond simply asking whether an agent can perform a task. They help answer a more important question: can the enterprise safely deploy autonomous agents at scale while maintaining control of its data?
The Future of Agentic AI Depends on Trust
The first chapter of enterprise AI was about what models could know. The next chapter is about what agents can do.
As organizations move from AI assistants to autonomous systems, trust becomes the defining challenge. The organizations that succeed will be the ones that combine intelligent agents with trusted storage infrastructure, resilient data platforms, and the ability to recover quickly when things don't go as planned.
The answer is not to slow innovation. The answer is to build infrastructure that can contain risk, govern access, protect data, provide visibility, and recover quickly when needed. In the age of agentic AI, trust starts with resilient storage infrastructure.
For a broader perspective on agent identity, governance, hybrid cloud, and enterprise trust in agentic AI, I encourage you to read Rob Thomas’s companion blog, which examines how IBM is helping clients build trusted agentic AI environments across the full stack.
https://newsroom.ibm.com/blog-building-trust-into-the-next-generation-of-ai-agents