Federal Agencies Rush AI Adoption Without Full Security Safeguards

Matt Parks, chief artificial intelligence officer for the U.S. Army Corps of Engineers, is pushing hard to deploy AI across the agency. But he’s making one crucial point: cybersecurity must come first, not later.

That tension sits at the heart of federal agencies’ race to adopt AI tools while facing a problem that won’t go away. The drive to modernize is real. So are the risks.

Parks oversees three strategic pillars: speeding AI adoption, educating and training the workforce, and building tested, trusted AI capabilities. He’s invested in modern infrastructure, resilient system design, data governance, and Army Vantage to put data to work. Here’s the uncomfortable truth: the Corps faces constant cyber attacks from what Parks calls “bad foreign actors” targeting critical infrastructure, including hydropower dams and locks. The agency is now exploring AI-powered threat detection to sort through thousands of cyber risk alerts, many of them false alarms, so staff can focus on genuine threats.

That’s a practical response to a real constraint. Federal agencies are understaffed. Their networks take a pounding daily. AI could help—but only if security comes first.

The Human Cost of Rushed Implementation

Rudy Rojas, an information technology manager at the Department of Agriculture, offers a model built around human oversight. The USDA’s cybersecurity and privacy operations center uses AI tools to monitor networks and analyze threats. But humans stay in control. When AI flags suspicious activity, Rojas reviews the reports and decides whether to whitelist or blacklist applications. “So far, it’s working good,” he said. The agency hasn’t handed decision-making over to machines.

The risk of skipping that step is tangible. Kevin Dorsey, director of federal solutions development at Optiv+Clearshark, points to a hard constraint: federal agencies have stretched budgets. GPU purchases cost a lot. Data center space is limited. Agencies can’t afford AI systems that fail, need constant retraining, or rely on external infrastructure that might vanish. “We don’t have the luxury of, maybe call it luxury, going to Deep Seek and pulling in a Chinese model with our customers,” Dorsey said, pointing to real concerns about foreign AI dependencies and national security.

Questions About Dependence and Control

At Idaho National Lab, cybersecurity leaders are asking harder questions about what happens when AI systems fail. Scott Bowman, technical lead for cyber-physical systems at INL, raised concerns about reliance on application programming interfaces and cloud APIs. What if they become unavailable? What if budget limits or token constraints hit? And what if an AI system becomes so embedded in critical operations that agencies can’t function without it?

Bowman frames the issue through three lenses: securing from AI, securing with AI, and securing AI itself. That third category gets too little attention. The lab examines AI tools’ code and architectures through what Bowman calls “mechanistic analysis”—trying to understand how AI works well enough to trust it. But he’s honest about the limits: “You’re never going to get comfortable, and if you do get comfortable, then it’s probably not the right direction.”

That isn’t paralysis. It’s appropriate caution.

The Practical Path Forward

Chris Cardran, a federal security advisor for defense and intelligence at Splunk, recommends what he calls a “smart machine model” (SMM) approach. These smaller, task-specific AI systems differ from massive large language models trained on broad data. SMMs need less computing power, can go offline for updates, and remain self-contained. They’re designed for what Cardran calls “machine-based data”—cyber logs, metrics, time series forecasting—not cognitive tasks that might require broader autonomy.

Operational technology systems are where AI could bring genuine benefits, Cardran argues. But that depends on building security architecture first and adding AI capabilities afterward.

Parks put it plainly: cybersecurity must be integrated into everything the Corps does. The AI ecosystem needs three things operating at “a high, high level”: capabilities, authoritative data, and modern IT infrastructure. All three need a foundation of strong security and human oversight.

This isn’t anti-AI sentiment. Federal agencies serve the public. Their systems protect critical infrastructure. Their networks hold sensitive data. Moving fast and breaking things isn’t an option.

Why This Matters:

Federal agencies control critical infrastructure—dams, power grids, water systems. Those systems face constant cyber attack. The temptation to deploy AI quickly to handle threats is strong. But hasty deployment without strong security architecture and human oversight could create new vulnerabilities instead of closing existing ones. The agencies described here are doing this work responsibly: building safeguards, keeping humans in control, and questioning dependencies. Still, pressure to speed AI adoption across government is real, and not all agencies have the resources or expertise that the Corps of Engineers, USDA, and INL bring to this challenge. Risks build up in the gap between best practices and actual deployment across the federal government. Public trust in government systems depends on agencies getting this balance right: advancing AI innovation while maintaining the security standards that protect shared infrastructure and public data.