Pentagon's AI Push Must Prioritize Cybersecurity, Officials Warn

The U.S. Army Corps of Engineers is building its artificial intelligence strategy around three pillars: accelerating AI adoption, educating and training its workforce, and developing AI capabilities that are tested, trusted and verified. But there’s a catch: cybersecurity must run through every one.

Matt Parks, the Corps’ chief artificial intelligence officer, stressed the point during a federal cybersecurity forum. “I’ve made it a real point to include cyber in everything that I do to make sure that the AI ecosystem triad of capabilities, authoritative data, and modern IT infrastructure are operating at a high, high level,” Parks said. The stakes are real. Federal agencies control critical infrastructure—dams, locks and power systems—and face relentless cyber attacks. The Corps can’t roll out AI tools without hardening them first.

The Infrastructure Challenge

The Corps already uses digital twins in an operational technology monitoring initiative to protect hydroelectric dams, locks and similar facilities. Parks described the threat plainly: “We do get constant cyber attacks from bad foreign actors on those things. So we’re always trying to mitigate that risk.” The agency is assessing AI-powered threat detection systems that could sort through thousands of cyber risk alerts, many of them false positives, leaving human analysts to focus on genuine threats.

That approach reflects a broader principle gaining traction across federal IT: AI works best when it amplifies human decision-making, not when it replaces it. At the Department of Agriculture, the cybersecurity and privacy operations center uses AI to monitor agency networks, analyze threats and produce centralized reporting. Rudy Rojas, an information technology manager and security specialist there, described how the system works. “These reports will go back to somebody like me. I’ll review those reports and determine whether we need to do a whitelist or whether we need to create a blacklist,” Rojas said. The USDA keeps people in the loop. “So far, it’s working good,” he noted.

The agency is also adding AI to its zero trust security architecture. Its chief information officer is prioritizing endpoint protections and software update management.

Avoiding the AI Dependency Trap

At the Idaho National Lab in the Energy Department, cybersecurity leaders are weighing risks that AI itself can create. Scott Bowman, the technical lead for cyber-physical systems at INL, frames the challenge in three parts: securing systems from AI threats, using AI tools to secure operations, and protecting the AI systems themselves.

Bowman’s team runs the COGNITO program, which assesses consequences before agencies adopt new AI systems. It examines how AI could affect existing workflows, which systems would depend on it, and how critical it might become. These aren’t abstract questions. Bowman raised a practical concern: what happens if an application programming interface, or API, or cloud API becomes unavailable? What if budget or token limits run out? Some mission-critical systems can’t tolerate that risk.

INL also conducts what Bowman calls “mechanistic analysis” of AI tools’ code and architectures to understand how they work at a granular level. Still, he’s clear-eyed about the limits. “You’re never going to get comfortable, and if you do get comfortable, then it’s probably not the right direction,” Bowman said. That’s pragmatism: recognize what you can and can’t know about complex systems before deploying them at scale.

Choosing the Right Tool for the Job

Chris Cardran, a federal security advisor for defense and intelligence at Splunk, recommends that federal agencies begin with smart machine models rather than large frontier AI models. SMMs are built for specific tasks, need less computing power and can remain self-contained, making them easier to update, test and redeploy. Cardran pointed to operational technology systems as a prime use case. “I think OT is probably one of the greatest places where you’re going to see some great benefit of AI,” he said.

Kevin Dorsey, the director of federal solutions development at Optiv+Clearshark, says the core challenge is finding practical AI applications that improve mission success while meeting government governance requirements. Budget constraints matter. Agencies have limited funding for GPU purchases and limited data center space, so they’re seeking tools that deliver results without requiring massive infrastructure investments. Dorsey said agencies need models they can run locally, rather than external services that depend on foreign or third-party infrastructure. “We don’t have the luxury of, maybe call it luxury, going to Deep Seek and pulling in a Chinese model with our customers,” he said.

The focus is narrowing to concrete gains: AI assistance in detecting adversary tactics, documenting procedures and collating operational intelligence. These applications are measurable and defensible, improving security without creating new dependencies.

Parks emphasized that success demands more than technology. “This work also requires a culture of cybersecurity awareness” across the organization. Systems are only as secure as the people who use them.

Why This Matters:

Government agencies control critical infrastructure that millions depend on. Rushing to adopt AI tools before ensuring they’re cybersecure could expose dams, power grids and water systems to attacks at scale. The officials quoted here advocate a disciplined approach: assess risks before deployment, keep humans in control of security decisions, choose smaller, purpose-built AI models over massive external systems, and recognize that no system will ever be perfectly secure. This is a market-aware, risk-conscious philosophy, much like the way private companies evaluate infrastructure investments. It rejects rushing new technology into production without vetting it first. The stakes are national security and public safety.