Presented by WGU


Most enterprise security teams cannot tell when an AI agent goes rogue, because few of them have measured what normal agent activity looks like on their own networks, Paul Bingham, senior vice president and executive dean of the School of Technology at Western Governors University (WGU), told VentureBeat editor in chief Matt Marshall at the recent VB AI Impact Series event.

Bingham, who spent 24 years as an FBI special agent including working cybersecurity investigations before coming to WGU, said companies also hand agents credentials before anyone decides what each agent may touch, so a misbehaving agent can reach far more than its task requires.

"I'm seeing ads on TV from well-known companies telling people my age to turn it all over to the agent, and it'll check your email and make your reservations," he said. "That's like giving your 10-year-old the keys to the car."

Enterprise leaders need to make three moves:

  • restrict what each agent can reach;

  • learn what normal agent activity looks like so deviations stand out;

  • and put the next security dollar into trained people before buying more tools.

Cybercriminals now use the same AI attack tools as nation-state hackers

High-profile breaches this year illustrate the range of threats security teams face from patient, state-backed attackers to smash-and-grab criminal groups. The October 29 Binance theft underscores the risks posed by compromised employee devices, while the widely reported July breach at Hugging Face highlights the emerging threat of AI agents penetrating production infrastructure.

The distinction between those attackers is blurring. Smash-and-grab criminals now wield nation-state tools, while state-backed attackers, once slow and stealthy, are picking up speed.

“The moral of the story is to protect against both, and just assume that you've been breached,” Bingham said.

For AI agents, that means being able to contain a problem quickly — and VB’s monthly Pulse survey suggests that how companies manage agent identities matters. The survey found that 46% of enterprises reported an agent incident or near miss in the past year. Among companies that give each agent its own identity, 60% said they could shut down a rogue agent within minutes, compared with 33% of companies whose agents run on shared keys.

But responsibility starts before an incident occurs. Too often, network administrators and security teams learn about the problem only after an agent has caused damage, leaving them to handle the cleanup. Or, in Bingham’s analogy, fix the garage door the 10-year-old backed through.

"Whoever creates the agent has the first responsibility to not hand the keys over, or, if they're going to, to restrict how the agent uses them, whether that's permission-based or time-based access," he added.

Agents should also not be allowed to clear multi-factor authentication (MFA) prompts on their own, because MFA exists to stop automated agents from logging into accounts and remains one of the few protections companies still have, he said.

"Though I may love my agents, and they are special, I would not want to hand over that capability," he explained. "Take the extra two seconds, whip your phone out and decide each time whether you want to authorize this."

Detecting a rogue AI agent requires a baseline of normal agent behavior

Security teams usually catch an intrusion because something deviates from a baseline of normal activity. The VB survey found that roughly 75% to 78% of companies say they have a solid baseline for normal network activity, but the figure drops to about 50% for agents operating inside the network.

"I think they're exaggerating," Bingham said. "Most of them really don't know what normal activity looks like."

The attack vectors remain familiar to Bingham, who entered the cybersecurity field around 2003. What has changed is the pace. His advice: start with the “jewelry box” — the assets attackers most want — and build monitoring and detection around them.

“Hunt every day for the breach because it’s likely there,” he said.

Security leaders should fund staff training before buying more cybersecurity tools

Security leaders should make room in this year’s budget for staff training, Bingham said. Enterprises have already bought too many tools, or aren’t getting their full value, because they lack enough trained people to use them effectively.

WGU's vice president of security, James Chandler, who protects a fully online network that almost 200,000 students use every day, told Bingham he wants staff who can adapt enterprise AI platforms to the attacks adversaries keep reworking with AI. Bingham said AI will take over much of the analytical work in security operations centers (SOCs), raising the question of how those staff analysts’ roles should evolve.

"You help them become more like engineers, because then they're actually tweaking and adjusting those tools, and they bring a higher skill set to interpreting what those tools produce," he said.

Companies should default to growing that talent internally, since current employees already know the company's priorities and systems, he added.

"You're just layering on, in most cases, an adjacent skill set," he said.

WGU adds bachelor’s degrees in AI engineering, AI infrastructure and operations, and AI security

WGU recently launched a bachelor's degree in AI engineering, and in January it will add bachelor's degrees in AI infrastructure and operations and in AI security. Each is an addition to the existing WGU programs in software engineering, cloud and network engineering, or cybersecurity.

Graduates fill roles at companies that cannot grow that talent internally, Bingham said, but the programs also serve a second group.

"Some folks that are already practitioners come and find resources like us to say, 'What's the extra layer I need to add on that's adjacent to the knowledge I already have?'" he said.

A forward-deployed security specialist should pair security and development teams on AI projects

Ship fast and worry about security later is no longer advisable, Bingham said. Cybersecurity professionals should be in the room when an AI system is first designed, with each side learning from the other. Security staff can borrow urgency from developers, and when he asks security practitioners what their main job is, he hears answers such as defending the network or enforcing policy.

"And I gently say, 'Actually, your main job is to ensure the business can keep doing business,'" Bingham said.

Developers, in turn, can learn the security habit of imagining how someone might use what they built in ways they never intended. The idea of putting both groups on the same team borrows from the forward-deployed engineer, a role companies across the industry are rushing to create.

"I wonder if we created a forward-deployed security specialist within a company," he said. "Just create a new role and have people from your security staff or your development staff apply for it. It brings some pizzazz to the role, and the key thing you're looking for there is collaboration."

Visit this page for information about WGU’s School of Technology programs.