If you thought morality, ethics and a broader understanding of context and what's right and wrong didn't matter all that much, well, you'd be very wrong. That's one of the lessons emerging from the rise of agentic AI. Tasks conducted with ruthless, clinical efficiency might achieve an outcome in record time, but that outcome could be Pyrrhic in the extreme.
That's the point of departure Pete Harteveld, Exabeam CEO takes when discussing the emerging requirement for behavioural analytics extending to our heartless new co-workers. "Sure, agents can be compromised by bad actors, just like a human can. And sure, they can also be malicious and doing something that we would deem non-permissible," he said.
The bigger and arguably more commonly-encountered issue, though, has less to do with espionage, legerdemain and the spooky shenanigans of hooded hackers, and more to do with the nature of machine intelligence. "The construct of an agent is that it wants to accomplish its task as quickly and efficiently as possible. It is pure and that purity is not balanced with morality. I have a moral compass, and think twice if accomplishing my task is harmful to someone else or some other application."
An agent, said Harteveld, has no such qualms. "That purity can make it malicious without intent."
Fans of the Mike Judge television show Silicon Valley will note this untrammeled efficiency in the exploits of programmer Guilfoyle. Tasked with finding and eliminating bugs in critical software, his AI agent 'Son of Anton' is given permission to modify code. Determining that the most efficient way to get rid of all bugs is eliminating all the software (an action Guilfoyle hails as 'technically and statistically correct'), Son of Anton does just that.
As fiction, it was funny in 2019, when the show arguably anticipated AI and even Mythos. It wasn't quite as amusing in April 2026 when a Claude-powered AI agent deleted a firm's entire database.
"I mean, an agent will figure out a way, and we've seen it. We've done our own internal tests, and you can read things that Anthropic has put out there. AI agents will figure out a way to do their jobs as efficiently and effectively as possible, and it will exploit any gap that you leave it to be able to achieve that outcome," said Harteveld. "That is why behaviour is so important, because when it deviates, that's the signal you must understand and act on."
If agents can be capricious and do the unexpected, anyone who's worked in cybersecurity (or any form of user support) knows that humans have set a very high bar for a combination of weird and random, as well as predictable and anticipated actions. That's given rise to behavioral analytics, a data-driven approach studying and understanding activity patterns. It's not just people, either. Devices, applications, servers, service accounts, and of course AI agents all leave digital footprints that can be analysed and interpreted.
Baselines of what 'normal' looks like for each entity, such as login times, applications accessed, data volumes transferred, or API call patterns, are established. Continuously monitoring flags deviations from the baseline which could indicate imminent threats, which can then be managed or simply shut down.
Harteveld said where traditional signature-based or rule-based security tools look for known attack patterns, behavioural analytics is proactive and adaptive, focusing on how people and systems actually behave rather than what attackers are known to do. "If you don't understand behaviour, then you have left yourself in a reactionary position," he said. "And the difference today is that agents move continuously at speeds far greater than humans. The stakes have gone up quite considerably, and bad things can be significantly more impactful in the agentic environment."
Asked if this emerging issue is recognised by CISOs, Harteveld said it generally is, as there is a pronounced push to take advantage AI agents. "There's interesting dynamics in organisations. I recently met with a global CISO who said senior executives are demanding agentic AI adoption. Her issue was 'I don't exactly know how am I going to secure all of this'. So people are thinking about the challenge, looking for guidance and advice. When they're grounded in the fact that whether human or non-human, digital identities operate in similar ways, that gives the basis to put a layered security posture in place."
The big lesson that Harteveld again stressed, is that machine or not, baselining and understanding 'normal behaviour' is the foundation. "Without that, you're going to be in a reactionary posture. And that's a bad place to be."