Securing Agentic AI: A Defense-in-Depth Architecture for Autonomous Agents
- As enterprises push autonomous artificial intelligence systems out of testing labs and into production environments, they face a category of risk that application-level controls were never built to...
- Traditional single-layer security approaches fall short when protecting complex agentic deployments.
- The infrastructure layer acts as the foundational root of trust, answering a fundamental operational question: who is operating in the environment?
As enterprises push autonomous artificial intelligence systems out of testing labs and into production environments, they face a category of risk that application-level controls were never built to handle, according to Oscar Wahlberg, senior director of product management at Nutanix.
Self-governing systems capable of reasoning, making standalone choices, and performing actions throughout a setting bring about a danger tier that software-level safeguards were never designed to hold.
Why traditional security models fail for autonomous agents
Traditional single-layer security approaches fall short when protecting complex agentic deployments. Relying solely on basic software guardrails to catch a malicious prompt does not prevent an active agent from misusing legitimate access privileges during a runtime loop, as noted by Wahlberg.
When an autonomous system gains execution privileges across a data center, security postures must scale into a comprehensive defense-in-depth architecture. This multi-layered strategy spans infrastructure, storage, compute, networking, and a governing control plane, with each layer targeting a distinct category of risk rather than duplicating identical controls across the stack.
Infrastructure layer: Establishing a hardware root of trust
The infrastructure layer acts as the foundational root of trust, answering a fundamental operational question: who is operating in the environment? Before an enterprise can trust what an agent executes, it must verify the integrity of the underlying hardware and runtime environment.
Securing this layer relies on platform attestation, secure boot, and confidential computing technologies to prevent unauthorized access both inside and outside the server. For regulated sectors such as financial services, this hardware-level isolation ensures AI production workloads stay within their assigned operational scope, mitigating supply chain compromises and runtime tampering risks.
Network layer: Governing dynamic agent communication
Modern agents generate high volumes of east-west traffic by querying external data sources, calling APIs, and spinning up secondary agents. This dynamic concurrency bypasses traditional static network configurations and can mask unauthorized data exfiltration or lateral movement.
AI agents ought to be viewed as a brand-new category of network identity, ensuring an individual agent communicates strictly with approved peers or designated data repositories where permission has been expressly granted.
Oscar Wahlberg, Nutanix
To address this traffic complexity, Nutanix developed the Agent Gateway as part of its Agentic AI solution. Operating alongside zero trust segmentation capabilities like Nutanix Flow and integrations with Cisco Secure AI Factory environments, the gateway enforces dynamic policies that restrict interactions between models, tools, and enterprise databases by default.
Control plane layer: Managing permissions and token governance
While infrastructure and networking secure the environment, the control plane serves as the operational brain, providing a centralized point to manage agent permissions, resource consumption, and tool access. Wahlberg points out that enterprises frequently underestimate this specific layer, which is crucial for managing Day 2 operations.
The centralized AI gateway enables IT administrators to audit model access, protect Model Context Protocol (MCP) tools, and prevent excessive token consumption caused by agents getting trapped in execution loops. By integrating hardware compute from Intel—such as Intel Xeon 6 processors with built-in AMX instructions for efficient inference—alongside Cisco network fabrics and Nutanix control software, organizations can deploy a unified, multi-vendor AI Cloud capable of scaling thousands of autonomous agents safely.
