The agentic frontier: A CIO’s guide to securing autonomous AI

While first-wave AI was largely conversational—chatbots that summarized documents or drafted emails—the second wave is operational. AI agents are autonomous entities capable of reasoning, planning, and executing multi-step workflows. They don’t just tell you that your inventory is low. They negotiate with suppliers, update ERP systems, and optimize shipping routes without human intervention.

However, this autonomy introduces a paradox. The more useful an agent becomes, the more dangerous it can be if compromised. While we can add humans-in-the-loop, when agents gain the power to act on behalf of the enterprise, they expand the attack surface beyond traditional security frameworks. In fact, current industry data suggests that nearly 90% of IT leaders have already experienced security incidents related to AI pilot programs.

To move to a production-ready AI factory, CIOs must move beyond wrapper security and embrace a framework of sovereign AI. HPE and NVIDIA have identified three ways to secure the agentic enterprise: Establishing a silicon-level root of trust, implementing zero-trust identity for non-human actors, and deploying real-time behavioral guardrails.

1. Building on a silicon root of trust

Security for autonomous agents cannot exist solely at the software or application layer. If the underlying infrastructure is compromised, every decision the agent makes—and every piece of data it touches—is at risk. In an era where model poisoning and firmware hijacking are real threats, security must be anchored in the hardware itself.

The vulnerability of “black box” infrastructure

Many organizations began their AI journeys in the public cloud. While convenient for training, the public cloud often limits visibility into the physical security of the stack. For AI agents handling sensitive intellectual property or regulated customer data, this lack of control is a liability.

The solution: Hardware-enforced integrity

The foundation of a secure AI agent is a silicon root of trust. Through HPE servers—specifically optimized for NVIDIA’s Blackwell architecture—security is embedded directly into the motherboard.

• Immutable fingerprints: Before the server even boots, the system verifies that the firmware has not been tampered with. If the “fingerprint” doesn’t match, the system simply will not start. This prevents low-level attacks that could otherwise allow a malicious actor to intercept data as it moves between the CPU and the GPU.
• Confidential computing: In high-stakes environments, data must be protected even while it is being processed. Confidential computing keeps the data being utilized by an AI agent encrypted in memory. This “trusted execution environment” (TEE) means that even if a system administrator or a sophisticated hacker gains access to the host operating system, they cannot see the data inside the GPU’s secure enclave.
• The sovereignty advantage: By deploying a sovereign AI factory, organizations can keep their models and agents on-premises or at the edge. This can include an air-gapped environment where the AI does not have to call home to a third-party provider, so the organization maintains sovereignty over its most valuable digital assets.

2. Zero trust governance for the “agentic identity”

The second major security hurdle for CIOs is the identity explosion. We are rapidly approaching a reality where there are more AI identities in a corporate network than human identities. Unlike human employees who have predictable working hours and clear organizational charts, AI agents can spin up thousands of sub-tasks and API calls in seconds.

The risk of over-privileged agents

Most AI agents are currently overprivileged. Developers often grant agents broad read/write access to databases so they just work. However, if an agent is compromised via a prompt injection attack—where a malicious user tricks the AI into ignoring its original instructions—that agent can become an internal threat, exfiltrating data or deleting critical records.

Implementing a “least privilege” framework

To secure these entities, CIOs must treat AI agents as high-privilege users, subject to the same zero-trust principles as any human executive.

• Dynamic authorization: Traditional role-based access control (RBAC) is too static for the speed of AI. Using NVIDIA’s AI-Q blueprint, enterprises can implement dynamic authorization. If an agent designed for marketing analytics suddenly attempts to access payroll data, the request is instantly flagged and blocked, regardless of the agent’s general credentials.
• Micro-segmentation for AI workflows: Organizations can use micro-segmentation to isolate agentic workflows. By creating digital blast cells, a CIO can ensure that even if one agent is compromised, it cannot move laterally across the network to infect other systems.
• Auditability: Every action an agent takes must be logged in a tamper-proof audit trail. Using integrated management capabilities in HPE and NVIDIA’s joint solutions, security teams can reconstruct the thought process of an AI agent. If an agent makes a faulty or dangerous decision, the CIO can trace it back to the specific data source or prompt that caused the deviation.

3. Real-time behavioral monitoring and guardrails

Unlike traditional software, which is deterministic (Input A always leads to Output B), AI agents are fluid. They reason their way toward a goal, and that reasoning can occasionally lead to hallucinations or jailbroken behavior.

Beyond static firewalls

A traditional firewall cannot stop an AI agent from accidentally leaking trade secrets during a negotiation. Securing agents requires a new category of active security that monitors the intent and output of the AI in real-time.

Set agent guardrails

As part of the HPE AI Factory ecosystem, developers can set agent guardrails.

1. Topical guardrails: Keeps the agent on task. If a procurement agent is asked about the company’s political stances, the guardrail intercepts the query before the large language model (LLM) even processes it.
2. Safety guardrails: Prevents the agent from executing dangerous commands, such as “Delete all records” or “Export user table to external IP.”
3. Output filtering: Real-time scanning of the agent’s responses to ensure that no personally identifiable information (PII) or proprietary code is being shared with unauthorized parties.

The strategic path forward: Building a secure AI factory

For the CIO, the goal isn’t just to “secure AI”—it’s to build a resilient, scalable, and secure AI factory that fuels business growth. Securing AI agents shouldn’t be seen as a bottleneck. It is actually the primary enabler of speed. When the business knows that its agents are running on a silicon root of trust, governed by zero trust identities, and protected by real-time guardrails, they can innovate with confidence.

The partnership between HPE and NVIDIA provides the full stack of HPE AI Factory solutions designed for this new reality. By combining HPE’s decades of experience in secure, hybrid infrastructure with NVIDIA’s world-leading AI software and hardware, we are giving CIOs the tools to lead the agentic revolution safely.

Conclusion: The 90-day mandate

The window for experimental AI is closing. As your competitors begin deploying autonomous agents into their supply chains, customer service, and R&D labs, the security of those agents will become your most significant competitive advantage.

In the next 90 days, CIOs should prioritize three actions:

• Audit the underlying hardware of existing AI pilots to ensure it supports confidential computing and silicon root of trust.
• Classify every AI agent as a unique identity with restricted, monitored access.
• Deploy a private cloud or on-premises environment for sensitive workflows to reclaim data sovereignty.

The future is agentic. Let’s make sure it’s secure. Learn more about the HPE Sovereign AI Factory.

*******************

As AI becomes increasingly central to economic competitiveness, scientific advancement, and national priorities, organizations require infrastructure that balances performance with security and sovereign control. Together, HPE and NVIDIA co-engineer rack-scale AI systems that integrate AI computing, high-performance networking, and supercomputing expertise to support large-scale AI workloads. This provides enterprises, governments, and research institutions with a trusted foundation for sovereign AI initiatives while maintaining control over critical data, models, and operations.