AI agents land on federal networks like new employees
Agencies are discovering that deploying AI demands the same governance as hiring talented staff, and infrastructure designed for human users was never built for machine-to-machine traffic.
TL;DR
Federal agencies deploying AI agents are discovering the technology behaves less like software and more like a talented new employee, Fortinet Federal CTO Felipe Fernandez said at Federal News Network's Cyber Leaders Exchange 2026. The distinction carries real operational weight: AI agents need access boundaries, activity monitoring, and ongoing supervision, governance work that falls on security teams already managing legacy systems and preparing for post-quantum cryptography. Fernandez said the shift demands operating through incidents rather than just preventing them, and warned that federal networks designed for human-to-application interaction cannot sustain the volume of machine-to-machine traffic AI workloads generate.
Felipe Fernandez, CTO at Fortinet Federal, offered a diagnosis at Federal News Network's Cyber Leaders Exchange 2026 that cuts against the way most agencies think about AI deployment. The technology doesn't behave like software, he argued. It behaves like a talented new employee, one that needs access controls, activity monitoring, and ongoing supervision. "It's not just a set-it-and-forget-it, fire-and-forget type tool," Fernandez said. "It is certainly like adding a very talented human to the environment." That reframing has practical consequences. When you deploy software, you configure it. When you hire someone, you govern them. The second model is substantially heavier.
And it lands on security teams that are already stretched.
The governance burden is only half the problem. Fernandez argued that agencies also need to abandon a mindset built around keeping adversaries out and replace it with one built around sustaining operations when something inevitably gets through. "We have to think about operating through incidents, not just recovering from them," he said. "We have to understand we're operating in an environment where we assume something bad has already happened, something bad is happening, something bad will happen." That is the resilience pitch, and it implies a different set of investments (continuity planning, redundancy, failover architectures) than the perimeter-defense model most federal networks inherited.
The infrastructure itself is a constraint.
Federal networks were designed for an era when users interacted with applications. Today the traffic runs in the other direction, applications talking to applications, automated workflows chaining across systems, AI agents reaching into data repositories. "The infrastructure wasn't going to be laid out forever, like roads and highways," Fernandez said. "It needs to be upgraded to support the additional traffic that's being put on it." The traffic itself has changed shape. Machine-to-machine communication generates volume and patterns that human-driven networks were never engineered to carry.
Compounding the problem: many of the same agencies are simultaneously preparing for post-quantum cryptography migration. Before replacing vulnerable encryption, agencies need to understand where that cryptography lives, inventorying dependencies and assessing whether aging hardware can handle modern algorithms. Fernandez said most organizations remain in the inventory phase. The AI governance problem and the PQC problem run on the same infrastructure, and neither tolerates delay.
Published ·Deep Fathom