Artificial intelligence is accelerating cybersecurity threats, while AI adoption is posing urgent risks around data, identity, trust and operational decision-making. Agencies grappling with those challenges are exploring how to securely adopt AI for both general purpose use and cyber defense.
Thomas Edgar, the chief cybersecurity scientist at Pacific Northwest National Laboratory, said PNNL’s research explores both the risks and opportunities posed by advanced AI models.
“Trying to benchmark and understand how bad these models could be if they’re used to target our critical infrastructure,” Edgar said during AI driven defense and the future of the SOC discussion. “But we do prototypes and proof of principle type research, and so AI has been a boon for us to help us kind of iterate and try things faster, supporting our reverse engineering, vulnerability analysis, protocol analysis, data analysis. We’re really looking at how we can really extract optimization out of AI to make us more efficient and effective, and then ultimately transition that off to our government sponsors and our critical infrastructure, so that they can take advantage of what we’ve learned.”
Organizations are adopting AI, especially for cyber defense, to help triage and analyze vast amounts of data. But a key piece of AI adoption is keeping a “human in the loop” to maintain control over critical decisions and actions.
Edgar said one facet of PNNL’s research is into “how to make sure that the human can take over and keep making good decisions when they’re not doing that level-one triage anymore.”
PNNL has also been exploring how operators can trust the decisions made by AI-enabled systems.
“We’ve been doing a lot of understandability into what the AI is doing, and trying to make sure that you have that provenance of its decisions, so that you can trust what it’s doing, and be okay with it taking these automated actions,” Edgar said.
Justin Ubert, director of cyber protection at the Transportation Department, said DoT has also been examining the “decomposition of processes” involved in adopting AI.
“When you decompose them, then you can determine where handoffs [to] human in the loop, if you will, or where you can have automation, where you can have AI as a focal point, an agent doing some task, and where you need a human to intercept, pull in and make a decision,” Ubert said.
Complex and sensitive workflows, such as cybersecurity tasks, might require more consistent check-ins compared to less sensitive tasks.
“Those seem like singular events, but really we’re talking about check in and check out of the process for the human as things happen, so they’re really a focal point of a process flow that includes automation and the agent or AI conducting some task,” Ubert said.
Foundational risks and new opportunities
The adoption of chatbots and AI agents presents both foundational risks, as well as new and novel risks. Ubert pointed to the need to rely on existing, fundamental processes to address risks like data leakage and inconsistencies in model outputs.
“We have to be careful of ensuring that fundamental processes are in place,” Ubert said. “We know how to vet and validate our data processes. So all of these things, just because we get to speed up certain things, should still be looked at and handled in appropriate fashion.”
But agentic AI, in particular, requires controls to address risks around “non-human identity sprawl,” Ubert said, as well as data access privileges.
“The agents and the model processes that are running all have an identity,” he said. “That identity is attributed to somebody who started that process, or at least controls that function, and that we can see the entire chain of whatever that agent is doing. But we also want to make sure we’re doing things like just-in-time permissions. So just as much permissions needed to do the thing, scoped very tightly to what is necessary.”
While AI has been able to accelerate the cyber attack cycle, Ubert said DoT is also examining how to leverage the technology to speed up defense.
“All of our tactics and techniques are predicated upon humans attacking, which have certain behaviors, certain methods, and these [models] don’t necessarily have to abide by those rule sets,” he said. “So we’re starting to look at things like some low-level automated threat hunting, some automatic quarantine-as-a-playbook. So if we determine an action is a high-risk action, doesn’t match our baseline, we can just stop it, quarantine it. If it is legitimate, somebody loses a few seconds at a time, and we can fine tune the model for that type of response again.”
Rob Sherry, a federal sales leader at Palo Alto Networks, said he’s seeing AI accelerate an ongoing shift in federal cybersecurity from traditional log collection and threat analysis to “real-time security operations.”
He pointed to the recent Office of Management and Budget memo that updated federal logging requirements to a risk-based, prioritized approach.
“That shift is moving traditional [Security Information and Event Management] and logging into real time security operations, getting into threat hunt, pulling in [Internet of Things] and OT,” Sherry said. “So we’re seeing before our very eyes — out of necessity, frankly, with the speed at which attackers can move — a fundamental change in how agencies, departments, are aligning themselves and the technologies that are going to hopefully keep them ahead of adversaries.”
Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.