Tri-Layered Security Framework for Agentic AI: A Robust Defense Architecture for Autonomous Systems

Sponsored by Nutanix

Autonomous systems capable of reasoning, decision-making, and action execution in various environments pose a new type of risk that typical application controls were not designed to manage. According to Oscar Wahlberg, senior director of product management at Nutanix, viewing this risk as a singular issue leads to incomplete architectural solutions.

“The protections meant to intercept harmful prompts fail to prevent an agent from misbehaving, potentially resulting in actions like unintentional database deletions or unauthorized data leaks through granted credentials that are improperly utilized,” Wahlberg states. “This highlights a core challenge as companies advance autonomous agents from trial phases to operational environments.”

When an autonomous system is given execution rights throughout a data center, it’s essential to implement a layered security strategy that covers all aspects of infrastructure, storage, computing, networking, and governance. Each layer should tackle a specific risk category instead of replicating controls across the entire architecture. No individual security measure or vendor can fully ensure protection; it requires a collaborative effort between all components.

By clearly delineating responsibilities across various layers and applying zero trust principles, organizations can establish a fortified security framework that enhances their overall defense. Recognizing which specific risks pertain to each layer transforms the concept of layered security into an actionable strategy, encompassing three levels, each with unique responsibilities.

Infrastructure Level: Building Trust for AI Operations

The primary role of the infrastructure layer is to establish a foundational trust, addressing the crucial question: who is functioning within the environment? Establishing a trusted identity is essential for all subsequent security measures. An organization must first trust the environment’s integrity before relying on an agent’s actions. When an agent seeks to perform an operation, verification is necessary to confirm that the request originates from the legitimate agent and not an impersonator.

Assuring this level of trust relies on technologies that anchor trust at the hardware level, such as platform attestation, confidential computing, and secure boot mechanisms, alongside safeguards against unauthorized access within and beyond the server. For highly regulated sectors like finance, this layer can isolate AI workloads to ensure neither the agent nor the operating environment exceeds its defined confines, thus mitigating risks like model tampering, supply chain vulnerabilities, and unauthorized access.

Network Layer: Managing AI Agent Communications

As agents begin exchanging information with other agents, APIs, applications, and enterprise systems, they create a dynamic communication environment that conventional static network configurations may struggle to oversee. An agent capable of unrestricted API calls, data queries, and creating additional agents generates a complex array of east-west traffic that can obscure potential lateral movement or data breaches without the proper security layers in place.

“AI agents should be treated as a new type of network identity, ensuring they can communicate with others or access data sources only where explicitly permitted,” Wahlberg suggests. “This requires transitioning from fixed static rules to dynamic policy enforcement.”

Nutanix’s Agent Gateway is part of its Agentic AI initiative, functioning as a cohesive governed framework to regulate costs and manage autonomous agent interactions. Combined with agents adhering to zero trust segmentation and utilizing features like Nutanix Flow for micro-segmentation—integrated with networking vendors such as Cisco Secure AI Factory—Agent Gateway facilitates effective governance over interactions among agents, models, data repositories, and enterprise applications.

This network layer controls lateral movements, data transfer, and regulates the interactions agents can have. Implementing a zero trust framework, with blocked access by default and scalable monitoring capabilities, is critical due to the unpredictable nature of agent behaviors. The Nutanix integration with Cisco UCS servers and Cisco AI PODs provides the necessary physical infrastructure (compute, storage, and networking) that supports AI operations.

Control Plane Layer: Regulating AI Agent Actions

The control plane serves as the command center, managing agent permissions, access to tools, resource utilization, and oversight during runtime. It’s vital to have a singular point for enforcing policies uniformly, rather than reinventing guidelines for each agent, according to Wahlberg.

“Agent Gateway operates as a universal endpoint for various models and tools, allowing IT teams to set their agents to communicate with this centralized control point,” he elaborates.

This centralized AI gateway enables administrators to monitor, audit, and control access to models and critical tools, safeguarding data and managing privileged access. This layer aims to mitigate risks such as misuse of privileges, unregulated agent activities, unauthorized tool access, data leaks, and excessive model use that may cause increased operational costs when agents become stuck in loops. Governance must be treated as a dynamic control system rather than an afterthought for compliance.

Why Uniform Security Approaches Fall Short in Agentic AI Environments

A significant architectural error organizations make is to assume that a uniform security model can fit every layer of an AI framework. Relying on application-level software for hardware-level trust or static legacy rules to manage dynamic agents can create systems that either hinder agent functionality or leave vulnerabilities unguarded. This one-size-fits-all mindset can lead to notable performance issues and operational challenges.

“Neglecting to designate specific roles to the appropriate layers results in governance blind spots,” Wahlberg explains. “Organizations might secure model outputs while overlooking data leaks between agents or restrict network access without having the control plane visibility to recognize if resources are being consumed inefficiently due to agents stuck in active loops.”

Concentration solely on the model could reveal the most significant vulnerabilities, as a guardrail meant to intercept harmful prompts does not prevent a misbehaving agent from misusing legitimate credentials. Incorporating security measures across the entire stack ensures that even if a model-level threat bypasses initial defenses, the agent remains restrained by trust established at the hardware level, network isolation, and access restrictions.

Collaboration of Intel, Cisco, and Nutanix in Layered Protection

The collaboration among these three companies showcases how the layered architecture effectively operates as a well-regulated, enterprise-scale AI cloud. Intel provides the computational power for agentic tasks and fortifies the execution environment via hardware-rooted trust and confidential computing while lowering costs through its processors. Intel’s Xeon 6 processors equipped with AMX enhance AI inference efficiency without relying solely on costly GPUs.

Cisco secures the environment, managing communications between agents and enterprise tools, while Nutanix supplies the software platform, minimizing architectural barriers and creating a central control plane that enforces permissions, visibility, and cost management, resulting in a cohesive defense-in-depth strategy for scaling autonomous AI operations.

Among the three layers, enterprises often undervalue the control plane the most, Wahlberg observes. A true control plane goes beyond initial setup, simplifying ongoing operations and providing IT teams with the insights and token management needed to keep autonomous agents secure and cost-efficient during production.

“In addition to choosing models and tools, the governance of agent deployments and their access to business tools within a tightly integrated platform will be vital for the success of AI initiatives,” he notes, predicting a future where organizations transition from a few AI applications to thousands of autonomous agents driving business strategies.

Leadership in technology should focus on creating a centralized governance layer now, capable of managing agent identities, tool permissions, and token allocations in real time, as this control point fosters the capability to scale securely.

“Establishing an AI system requires navigating numerous complex decisions,” he clarifies. “A control plane that communicates across different vendors and infrastructures is essential for implementing effective defense-in-depth strategies.”

Sponsored content is created by a company that has a business relationship with VentureBeat. For more information, contact sales@venturebeat.com.