The U.S. national labs are tasked with some of most challenging – if lesser-known – endeavors undergoing high-stakes innovation in the nation. The work requires a laser focus among the scientists, researchers and specialists engineering the game-changers and solutions for some of the most serious threats we face today and tomorrow.
In these scenarios, it’s all hands on deck. Everyone plays a part in expediting the proverbial key to unlock the future, whichever angle of the future that may be. It’s serious work that demands serious support.
Idaho National Lab CIO Mark Holterman is all in, as is his IT team and their partners.
“The key is creating those partnerships with the mission teams, ensuring they have everything they need to get their problems solved, get their challenges looked at – take away the burden, if you will, of operationalizing some of their proofs of concept, some of their research,” Holterman said. “On the operations side, keeping things running 24-7 flawlessly and securely is the name of the day. And we operate at that that pace.”
Holterman is perhaps the first to say that pace of operations is quickly reaching the limits of human-driven IT management. He’s bullish on the incorporation of autonomous AI to synthesize the growing volumes of data faster, more efficiently and more effectively.
Defining autonomy, establishing common ground
He’s also quick to describe what he means by “autonomous” – and what he doesn’t mean. Human checkpoints, reviews and validation are requisites. Fully autonomous? Not quite.
“Automation has been around for, you know, a while. However, the tooling, the compute behind the tooling and the availability of the tools themselves have advanced dramatically in the last few years,” Holterman said. “In the federal context, in the lab context – and I would say this isn’t terribly different than the commercial space, depending on the nature of the industry – we have to validate the results. But there’s tremendous amount of power in being able to look at something that’s rules-based, deterministic, script-driven type of automation.”
Agentic AI, which he characterizes as basically automations triggering other automations and multiple automations working at once, creating a high confidence an agent could be created to perform the human checkpoint role. While technically possible, it’s not acceptable, he noted.
“There still needs to be these checkpoints. In our context, obviously we need to make sure everything’s auditable, traceable, defendable. You have to have good logging throughout,” Holterman said. “But it is definitely changing the pace at which we’re able to get things done.”
There may be no clearer real-life example of this than the work the lab is doing to protect critical infrastructure – in some ways an extension of its cybersecurity work, but not quite the same.
“We research protecting critical infrastructure, and the amount of data that you can get from sensors and logging of different types of processes and such, the ability to mine that data and detect, preempt and prevent something from occurring is something we also on the traditional cybersecurity side. So it’s helping us with defense, while we also need to make sure that we’re having that human in the loop,” he said.
That defense is increasingly critical. Malicious actors are increasingly building agentic tools that comprehensively seek flaws and vulnerabilities in U.S. systems, accelerating the AI race and the lab team’s pace.
“We’re trying to be ahead of [zero days] to determine where we’re seeing that behavior with large models themselves analyzing the data, things that would take multiple analysts. You can have the agent surface these things faster and detect and prevent and act,” Holterman said. “That’s the next piece: It’s not only just providing an alert but taking some action. It could be as simple as quarantining, or it could be as dramatic as shutting down access to an unknown entity or something that’s causing a problem.”
Instituting the guardrails: where autonomy stops and human judgment starts
As Holterman pursues an admittedly aggressive path automated AI, he’s ensuring he’s doing so with the inclusion of other experts, researchers, participants and the larger AI community.
That includes a Center of Excellence for advanced analytics, machine learning and LLM models for working with large quantities of data. There’s also a Center of Excellence around intelligent automation – exploring everything from simple automation to the more advanced version with AI and a queue of work. An entire cross-lab group votes and prioritizes what gets done, how and when.
“We’ve created a ‘citizen development’ type of governance structure. It’s a Community of Practice. We have guardrails drafted, written out, there to have folks stay within some boundaries. The auditing of some of that is tough,” he said. “We want the power and the momentum and the efficiency and productivity that can be gained out of these quick wins but what’s tough for folks to know is when they’ve crossed certain lines.”
The citizen-development community fosters discussion, problem-solving and collaborates – even votes – on guidance. It’s still a precarious balance, but Holterman believes it’s essential to moving the needle.
“As much as I’d like to be conservative from the security perspective, we want to advance the mission,” he said. “So we’ve got to keep these things moving.”
Copyright
© 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.
