Fighting AI with AI requires enduring, new approaches

Watch the full discussion here.

The expansion of artificial intelligence capabilities has changed the way agencies work. From the straightforward use of chatbots to the advanced tools like agentic agents, the Trump administration is pushing hard on agencies to use AI to improve how they deliver on their missions.

At the same time, agencies have to keep up with how to govern and manage these models as the speed of release increases. And then there is the security side. Not only do agencies need to secure the tools they are using, but they have to defend networks against adversaries using AI-fueled cyber attacks.

Landon Shaw, the senior architect for the Administrative Office of the U.S. Courts, said two advancements are driving the use of AI tools to secure systems and networks.

First, agencies have done a better job of breaking down data silos, leading to a better understanding of cyber threats and risks.

Second, Shaw said, is the rapid evolution of using multiple AI agents that work together to protect federal systems from cyber threats.

“The biggest win for people today is to not have disjointed repositories for these different sources of data. We’re seeing an advantage of taking all these different separate data source repositories for whatever type of threat intelligence data is being gathered and concatenating them into a single place, so that you can then also utilize machine learning or AI tools to get a broader picture of what’s happening in your organization, so instead of trying to go and do separate searches in different databases, you can see a progression in a single location,” Shaw said on the discussion Defense Reimagined: Cybersecurity in the Age of Intelligent Adversaries. “We’ve spent a lot of time and effort over the last several years to make this reality where we don’t have these separate silos. From what I’ve seen, this has been the biggest change in the way we operate, giving us a single pane of glass visibility into how each of these individual systems is operating. That has been a game changer because that allows you to very quickly plug in AI tools that can find that needle in the haystack and create a case file that isn’t just a single item or a single system’s logs, but this case file can be a combination of 25 different data sets from 25 different security products that together can be put into a case that can be either assigned to another AI or assigned to a human to investigate, and it can also detect threats before they happen.”

Sharper sword or more solid shield

That ability to bring threat intelligence data together has led to the ability to apply AI agents and subagents, especially as cyber attacks have grown more sophisticated and volumetric attacks can occur faster.

“Now with the ability to quickly deploy agents and subagents with frameworks like what came out from Open Claw, you can spin up subagents and you now have the ability to do this ‘if this, then that’ logic, where you can have a AI look at something and make a decision, and before that decision is implemented, you can have a subagent, another AI, check the first AI’s work. Then you can have a third agent look at your rules that you defined and say, ‘Okay, does this decision require a human in the loop, or can I implement this decision on my own?’” Shaw said. “I think this is a big distinction because now we’re not just talking about a single AI where you have to put all of your logic into and let this one AI make decisions. The last four months have changed the world with the ability to deploy rapidly, deploy agents and subagents, and you’re using things like Claude for different workflows.”

The ability to use AI agents and subagents is giving agencies what Brent Hansen, the chief technology officer at Optiv + ClearShark, called either “a sharper sword or a more solid shield.”

Hansen said when organizations are using AI to fight against AI-fueled attacks, it’s less about better prompt engineering and all about speed.

“Being able to do the vulnerability scanning, being able to move quicker than they’ve ever been able to package and exploit vulnerabilities is challenging. But on the flip side, you’re also equipped with the same toolbox and with the same expertise to be able to move quicker to plug those holes, but it’s a very delicate balance of what we’re empowering and what is out there on the streets,” Hansen said. “It’s really the continuous monitoring, evaluation, red teaming of the existing models that you have in place, so that when you do earn trust, and that trust happens either out on the edge or whether it’s something internal, that is a continuous responsibility.”

This trust of AI tools and capabilities also comes from creating a data fabric that cuts across the organization.

Hansen said this fabric becomes especially important as agencies push data to the edge.

“It’s all about the access controls and the observability to understand when it’s under attack,” he said. “We’re better than ever at collecting data, and so with that ability, AI is able to harness that universal language to interpret lots of unstructured data, where you don’t have to have everything in a very specific system log template. So that’s great, we can move much quicker. When you’ve got that universal data fabric, you’ve got your ear to the ground, and you’re able to make those detections much quicker, do better incident response, and then how do you do the enrichment of that data, so that in the rearview mirror, when there might be an incident, that you can have as much context about that particular event?”

Cyber battlefield is changing

The need to push data to the edge and collect data at the edge means the use of a zero trust architecture to protect that information becomes even more essential.

Kevin Walsh, the director of Information Technology and Cybersecurity for the Government Accountability Office, said with how fast data and attacks are moving, it’s more difficult than ever to keep the human in the loop.

“In that cyber battlefield, the human isn’t going to have as much of a role as they do today. The only role we’re going to play is in the predictive or the proactive, like setting up the guardrails and trying to be as thoughtful as we can for what that might look like, and how we want to put controls around those agents and those subagents all the way down the stack,” Walsh said. “We need zero trust to make sure that those agents are acting in our organization’s interest, and making sure that when they are taking those actions, sometimes we do want to give them elevated privileges, sometimes we need that to enact the appropriate defensive countermeasures, but we want it to do it thoughtfully.”

Walsh added that makes the training of the employees managing and monitoring the AI agents so much more important.

“They have to make sure you have your multi-factor authentication set up, and all of those predictive or the preparatory steps that the humans can take, including locking down your application programming interfaces (APIs), making sure that your people are not just tongue-in-cheeking the cybersecurity training, but you are actually testing and maybe red teaming your own networks and employees to make sure that they are capable of doing the right thing,” he said. “I think that the human is our role is going to be changing, especially when we’re talking about cybersecurity operating at machine speed. We’re going to be more in a supervisory or managerial role, trying to find, which is the real attack, which needle in this haystack, or even needle in a stack of needles, is the real one we need to care about, and the real one we need to act upon, especially when we’re talking about the volume that can come at us, the complexity, the battlefield is changing. We need to change with.”

Shaw said agencies have to continuously monitor and evaluate the AI agents, especially when the frontier models are changing so often.

Models drift over time

He said each agent needs its own identity that includes a small subset of credentials that allows them to do what they need to do.

“You’re going to monitor the AI agent, and you’re going to do the trust but verify, and then once they build enough trust that they’re not making ridiculous decisions, at that point you may bestow additional work on them or give them additional responsibilities,” Shaw said. “An example would be to give the AI very granular rights, such as the ability to quarantine or to lock an account, but not unlock an account. For instance, the AI could log into a system if it thinks that there’s an account that’s been compromised and put a lock on that account, maybe that’s all it can do. I think it’s going to need to be the trust but verify approach, where you give very small access to these agents, monitor them, look at their thinking, make sure they’re behaving correctly, and make sure you don’t allow them to do any more than what you’re okay with them breaking.”

Walsh added like with most technologies, you can’t “set it and forget it,” as models do drift over time.

Hansen said as AI continues to advance, agencies can’t overlook the importance of due diligence and governance.

“It also comes with the responsibility of ensuring that we’re following zero trust principles, that we’re still applying the basics around multi-factor authentication, data encryption, so keeping those basics, but also evolving them to meet the standards and the requirements of where AI is taking us and, probably more importantly, where we will take AI,” he said. “When you’ve got the human in the loop, just keep that skepticism and ensure that we are using critical thinking as we’re evaluating things before they are out in production, that we’ve done everything possible to ensure it’s equipped and ready.”

Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.