Every data center cybersecurity team confronts the same challenging dilemma: host-based agents use up CPU cycles that high-performance computing desperately needs. For years, the industry has attempted to find a middle ground. The more security measures you add, the more performance takes a hit; yet, the more you prioritize performance, the more vulnerable blind spots become.
Consider a clear example of such a blind spot: the space between a virtual machine (VM) and its underlying physical host. In March 2025, Broadcom addressed several VMware ESXi zero-day flaws that could break out of the VM sandbox completely. In 2023, the ESXiArgs campaign impacted roughly 3,800 servers worldwide.
In both cases, a single breach took down or locked up dozens of VMs at once. Host-based agents offered no protection because the attack targeted the hypervisor itself.
The answer isn’t fine-tuning; it demands a complete architectural overhaul by taking security off the host entirely. Data processing units (DPUs), embedded in each server, make this possible.
Running security tasks on the DPU rather than the CPU preserves host CPU and GPU resources for the workloads they were designed to handle. Even better, the DPU remains hidden and unreachable by attackers since it functions independently from the host operating system.
The outcome is tamper-resistant security, enforced at full network speed – with zero performance penalty.
Legacy Risks at a Modern Pace
Data centers have consistently ranked among the most difficult environments to protect. Physical servers run hypervisors. Hypervisors run VMs. VMs run containers. Every layer introduces abstraction, and every abstraction creates blind spots where assets slip out of management and vulnerabilities stay hidden.
Misconfigurations build up gradually. VMs are cloned from stale templates. Firewall rules pile up exceptions that nobody reviews. Servers keep running for projects that ended long ago because nobody wants to risk taking them offline.
Peripheral security offers limited help in these settings. Firewalls and network security tools watch north-south traffic (data entering and leaving the data center). But most data center traffic moves east-west (lateral communication between VMs).
Once an attacker gets past a single entry point, perimeter defenses lose all sight of what happens next. This is where attackers linger undetected, escalating privileges well beyond the reach of conventional network boundaries.
AI data centers carry all of these same risks and then multiply them at breakthrough speed. Short-lived network connections exist for hours (or mere minutes) before vanishing. VMs are spun up and shut down for specific tasks. Containers are managed across nodes that shift resources on the fly. These on-demand assets appear and disappear faster than any manual process or scheduled scan can follow.
When a single GPU cluster can be worth millions of dollars and every fraction of efficiency gained means a direct competitive edge, deploying host-based security agents simply isn’t practical. Regrettably, some operators are quietly turning off security on their most vital compute nodes and crossing their fingers that the perimeter holds. The math simply doesn’t work.
A Blueprint for a Better Tomorrow
Moving security from CPU-based agents to a DPU-driven architecture removes the security versus productivity conflict by placing the entire security stack onto purpose-built hardware. The DPU acts as a built-in sensor in each server, streaming telemetry and watching network traffic without any effect on host operations.
The performance gains are substantial. Always-on real-time monitoring through a DPU can outpace CPU-based methods – and speed is only part of the benefit. The physical separation between the DPU and the host enables zero trust enforcement at the hardware layer.
The DPU sits between the host and the network, applying zero trust to both. Every packet, every access attempt, and every process undergoes inspection and policy checks. Even if the host OS is fully compromised, the DPU’s hardware-level isolation preserves its authority.
From a visibility standpoint, a DPU-based approach delivers continuous monitoring across both physical and virtual infrastructure, covering east-west (internal) traffic and north-south (external) traffic alike. Deep packet inspection happens right at the endpoint, removing the need to route traffic through external appliances and eliminating associated bottlenecks.
At the same time, privacy safeguards are woven into the architecture. Only kernel-level structures and system metadata are collected – never user data or application-layer content. The result is full visibility without compromising sensitive information.
Enabling Security and Performance
For twenty years, data center security has been framed as an impossible choice: security or productivity. DPU-based security resolves that equation. For AI data centers, where the consequences are the greatest and performance demands are the most rigid, security and performance no longer have to compete.
Related: Cisco Patches Critical Vulnerability in Data Center Management Product
