By admin 
The advent of autonomous systems capable of independent reasoning, decision-making, and action execution across complex environments introduces a novel and significant category of risk that traditional application-level security controls were fundamentally not designed to address. Treating this emergent risk as a monolithic problem leads to incomplete and vulnerable architectural solutions, according to Oscar Wahlberg, senior director of product management at Nutanix. "The guardrails designed to catch a malicious prompt won’t stop an agent from hallucinating and doing something it never should have done, like accidentally deleting databases or leaking sensitive data with a credential it was granted but then uses for something entirely different," Wahlberg explains. "That’s the central problem as enterprises move autonomous agents out of experimentation and into production." This shift necessitates a paradigm change in how security is conceived and implemented, moving from perimeter-based defenses to a comprehensive, layered approach that addresses the unique challenges posed by intelligent, self-directed systems.
Once an agentic system is granted execution privileges across the entire data center, the security posture must scale proportionally, adopting a robust defense-in-depth architecture that spans infrastructure, storage, compute, networking, and a governing control plane. Each of these layers is designed to address a distinct category of risk, rather than attempting to duplicate the same controls across the entire stack. The fundamental principle is that no single security control or vendor can provide comprehensive protection on its own; true security in this new paradigm relies on the synergistic collaboration of these distinct layers. By meticulously dividing responsibilities across the stack and adhering to the principles of zero trust segmentation, organizations can forge a secure framework that significantly enhances their overall security posture. Understanding which specific risks belong to each layer is the key to transforming the abstract principle of defense-in-depth into a practical, actionable security framework. This framework is characterized by three distinct layers, each with a clearly defined and critical responsibility.
The Infrastructure Layer: Establishing Foundational Trust for AI Agent Operations
At the very foundation of this defense-in-depth strategy lies the infrastructure layer, whose paramount responsibility is to establish a root of trust. This foundational element answers a critical, albeit simple, question: "Who is operating within the environment?" This trusted identity serves as the indispensable prerequisite for every subsequent security control implemented at higher levels of the architecture. Before an organization can place any confidence in the actions performed by an AI agent, it must first be assured of the integrity of the environment in which that agent is executing. When an agent requests permission to perform an operation, the system must possess the capability to unequivocally verify that the request originated from the legitimate agent itself, and not from an entity attempting to impersonate it.
Delivering this level of assurance hinges on the deployment of technologies that firmly root trust in the hardware itself. This includes sophisticated mechanisms such as platform attestation, confidential computing, and secure boot processes. These are complemented by robust controls designed to prevent unauthorized access, both internally within a server and externally across the network. For industries operating under stringent regulatory requirements, such as financial services, this infrastructure layer provides the crucial ability to isolate AI production workloads. This isolation ensures that neither the agent nor the underlying environment can operate outside its precisely assigned scope, effectively mitigating a range of critical risks. These include model and runtime tampering, the insidious threat of supply chain compromise, and unauthorized access to highly sensitive AI workloads. The integrity of the underlying hardware and the immutability of the execution environment are therefore paramount to building a secure foundation for AI operations.
The Network Layer: Orchestrating Secure Communication for AI Agents
As AI agents begin to interact and communicate with each other, with various APIs, with enterprise applications, and with other critical systems, they generate a level of concurrency and dynamic communication that traditional, static network configurations were never equipped to handle. An agent configured to call APIs, query diverse data sources, and autonomously spin up additional agents without any inherent constraints can quickly create a sprawling, unmanageable web of east-west traffic. This complexity can easily mask malicious activities such as lateral movement by attackers or the covert exfiltration of sensitive data, particularly when robust network security layers are absent.
"We should treat AI agents as a new class of network identity, and make sure that an agent can only talk to other agents or data sources where it’s explicitly allowed to do so," emphasizes Wahlberg. "That means moving away from rigid static rules toward dynamic policy enforcement." This shift in thinking is crucial; static rules, designed for predictable, human-driven interactions, are ill-suited for the fluid and often unpredictable behavior of autonomous agents. The solution lies in embracing dynamic policy enforcement, which allows security measures to adapt in real-time to the evolving interactions of these agents.
Nutanix’s innovative approach addresses this challenge with its Agent Gateway, a core component of the Nutanix Agentic AI solution. This unified, governed layer is specifically engineered to provide essential cost control and governance capabilities, thereby empowering organizations to effectively manage autonomous agent users. When coupled with agents that are fundamentally grounded in zero trust segmentation principles and leverage capabilities like Nutanix Flow for microsegmentation, and when integrated with leading networking vendors such as Cisco, through its integration into the Cisco Secure AI Factory, Agent Gateway plays a pivotal role. It assists enterprises in governing the complex web of interactions that occur across agents, models, data sources, and enterprise applications. The network layer, therefore, becomes the critical gatekeeper, governing lateral movement, preventing data exfiltration, and meticulously controlling the agent’s network interactions. A zero trust framework, characterized by access being blocked by default and scalable interaction monitoring, is particularly important for agents, given their potential for exhibiting unreliable or unexpected behavior. The seamless software integration between Nutanix and Cisco UCS servers, as well as Cisco AI PODs, delivers the turnkey physical infrastructure – encompassing compute, storage, and networking – upon which these sophisticated AI factories can reliably run.
The Control Plane Layer: Governing Agent Permissions and Actions
The control plane represents the central nervous system of this entire autonomous AI ecosystem, acting as the ultimate point of management for agent permissions, access to tools, resource consumption, and runtime visibility. "What matters most is having a single place where policies can be enforced consistently rather than reinvented for every agent," states Wahlberg. This centralization is key to avoiding the fragmentation and inconsistencies that plague decentralized security models.
"Agent Gateway acts as a universal endpoint for different models and tools, so an IT team can configure their agents to talk to this single control point," he elaborates. This universal endpoint simplifies management and ensures a unified approach to security policy enforcement. The centralized AI gateway empowers administrators to meticulously observe, audit, and control access to AI models, as well as to the various tools that agents might utilize. This includes the implementation of robust mechanisms for protecting sensitive data and gating privileged access. This layer is specifically designed to mitigate a spectrum of critical risks, including privilege misuse, runaway agents that operate outside their intended parameters, unauthorized tool usage, data leakage, and the excessive consumption of computational resources—particularly token consumption—that can occur when agents become trapped in inefficient runtime loops. The effectiveness of this layer hinges on treating governance not as a mere compliance afterthought, but as an integral, runtime control system that actively shapes and constrains agent behavior.
The Failure of One-Size-Fits-All Security in Agentic AI Environments
The most significant architectural misstep that enterprises often make when adopting autonomous AI is the assumption that a single, overarching security model can be stretched to cover every layer of the complex AI stack. When an organization attempts to address hardware-level trust requirements with application-level software, or relies on outdated, static network rules to manage the dynamic and fluid nature of AI agents, it invariably constructs an architecture that is either overly restrictive, preventing the agentic system from performing its intended functions, or dangerously permissive, leaving critical security doors wide open. This "one-size-fits-all" approach not only leads to significant performance penalties but also introduces substantial operational friction, hindering the efficient deployment and management of AI initiatives.
"By failing to assign specific responsibilities to the appropriate layers, enterprises end up with blind spots in governance," Wahlberg warns. "They might secure the model output but miss that there’s data leakage between agents, or they might secure the network but lack the control plane visibility to understand that they’re wildly burning tokens because the agents are stuck in some kind of runtime loop." This highlights the critical need for a granular, layered security approach. Focusing exclusively on the AI model itself leaves the largest security gaps of all. A sophisticated guardrail designed to catch a malicious prompt, while valuable, is entirely ineffective at preventing a hallucinating agent from misusing a legitimate credential it has been granted. Embedding security principles across the full stack, from the hardware up to the control plane, ensures that even if a threat at the model level manages to slip past initial filters, the agent remains constrained by hardware-rooted trust mechanisms, network isolation, and granular access controls at the agent layer.
Synergistic Defense-in-Depth: The Intel, Cisco, and Nutanix Collaboration
The powerful three-way partnership between Intel, Cisco, and Nutanix exemplifies how a layered architecture can be practically realized to create a well-governed, enterprise-grade AI Cloud. Intel provides the essential computing power required to run demanding agentic workloads, securing the execution environment through its hardware-rooted trust technologies and confidential computing capabilities. Furthermore, Intel is actively driving down costs through its innovative accelerators, such as Intel Xeon 6 processors featuring built-in AMX, which efficiently accelerate AI inference without an exclusive reliance on expensive GPUs.
Cisco complements this foundation by wrapping the entire environment in a secure fabric that meticulously governs communication between agents and enterprise tools. This ensures that interactions are authorized and monitored. Nutanix, in turn, delivers the overarching software platform, crucially minimizing architectural silos and providing the central control plane. This control plane is responsible for enforcing permissions, delivering vital visibility into operations, ensuring cost governance, and ultimately, tying the entire architecture together into a cohesive defense-in-depth solution. This integrated approach empowers enterprises to scale their agentic AI deployments with confidence and security.
Of these three critical layers, Wahlberg observes that enterprises currently tend to underestimate the importance of the control plane the most. He stresses that a truly effective control plane extends far beyond the initial deployment phase, playing a vital role in simplifying Day 2 operations. It provides IT teams with the continuous observability and strict token governance that are absolutely essential to keeping autonomous agents secure and cost-effective in a production environment.
"Apart from model and tool selection, governing the agent deployments and their access to models and business tools in a tightly integrated full stack platform will be important for the success of AI projects," Wahlberg asserts, looking towards a near future where organizations will transition from managing a handful of AI use cases to orchestrating thousands of agents working autonomously to drive business objectives. Technology leaders must prioritize the development of a centralized governance layer today that possesses the capability to manage agent identities, tool permissions, and token budgets in real time. This control point is the fundamental building block that establishes the operational muscle required to scale AI initiatives safely and effectively.
"You can’t build an AI system without getting into a lot of complex decisions," Wahlberg concludes. "And you need a control plane that talks across multiple vendors and infrastructures to help you solve for those defense-in-depth strategies." This integrated, multi-vendor approach is not just beneficial; it is becoming an absolute necessity for navigating the complex security landscape of autonomous AI.
Learn more about the Nutanix Agentic AI solution here.
Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact [email protected].
