Artificial intelligence must remain traceable and accountable as it takes on more responsibility, particularly in cybersecurity applications where an algorithm’s mistake could lead to disastrous consequences, advises Pete Harris, head of cyber operations at the technology consulting firm Revolutional.
“We’ve gone into this AI world of, ‘We just want everybody to lean into it.’ … I think we’re learning some of the downfalls of that,” Harris said during Federal News Network’s Cyber Leaders Exchange 2026.
“We also have to have the policies and the regulations and accountability attached to that, to start to minimize some of those things and let chief information security officers sleep at night.”
Federal employees most commonly use AI as an assistant for everyday tasks — crunching data, generating reports and other efficiencies, Harris said. Some have begun experimenting with more complex AI tools to augment their work.
AI bots are proving particularly successful at niche work, like collecting a massive volume of threat intelligence, activity logs and data analysis in a security operations center, he said. Now, the challenge is deciding how a system should react to that information and whether human analysts should be involved in the response.
Accountability in the AI age
CISOs are learning to navigate the opportunities and challenges of unfamiliar AI in their systems. They’re wary of algorithms that behave like inexperienced employees, who may not understand the context of their work or the policies that govern it, and who can be easily manipulated, Harris said.
And CISOs are concerned about programs that want to please the user, even if it means making up an answer.
If organizations improve at checking how well a tool works, Harris said, they’ll become more skilled at training the algorithm itself and learning to spot hallucinations.
He argues cyber analysts will eventually take on the role of “AI commander” while the software sifts through the data.
“You’re going to see analysts going, ‘I’m in charge of 8,000 AI bots that are out there collecting data, and through that, I picked three that I went down and audited … and here’s the audit trail that you can follow of what those agents did,’ ” Harris predicted.
But that’s only for the AIs a CISO knows are there. Another worry, Harris said, is the bots that are part of a network without the organization’s explicit approval.
For instance, an app that a workplace adopted 10 years ago may now have an AI-powered chatbot with access to the organization’s data, he said. It’s also difficult to track how employees use publicly available options like Microsoft’s Copilot, Anthropic’s Claude or OpenAI’s ChatGPT, or which other AI applications workers may have on personal devices while working in the office or remotely.
“Now, you have nondevelopers able to do some development work and create workflows. Who’s double-checking those? Who’s validating those? Who’s approving those? And then when it doesn’t work, who’s making sure that it’s all cleaned up and all the data is sanitized?” Harris added. “That’s where you’re also seeing this uptick in shadow AI within organizations.”
He said there’s no silver bullet that can solve that problem. Employers could crack down on hidden AI tools by rescinding “bring your own device” policies and requiring staff to use work-issued computers — but good luck trying to take that flexibility away from employees now, Harris said.
Vetting AI vendors
To protect against breaches and leaks, federal agencies may be tempted to buy AI-enabled cybersecurity products that are too good to be true, he added. He warned government customers to be wary of sales pitches for fully autonomous network security products, including those that promise red-teaming and penetration-testing capabilities.
“They’re selling you a demo,” Harris said. “The technology is evolving fast, and these companies are doing some amazing things with AI, but we’re not there yet.”
He encourages those who are in the market for AI tools to ask vendors what the tool does when it fails, hallucinates or creates a bad outcome. Ask: How can the user track and audit its work? How do the ripple effects spread, and how can an organization minimize harm?
“Don’t just show me that it’s got logs,” Harris said. “If the AI, at 2 a.m., detects something and takes a reaction and it’s wrong, what’s the outcome of that?”
But he downplayed concerns that the technology presents unique security challenges or could put scores of employees out of work, suggesting tried-and-true protective measures like authority to operate and thorough training still apply.
“We just look at AI like it’s something different,” he said. “It’s really not.”
Discover more articles and videos now on our Cyber Leaders Exchange 2026 event page.
Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.