Behavioral Security Is Mainstream; the Hard Part Comes Next

Behavioral security is firmly established as part of the modern security stack—and the trajectory is still upward.

“Behavioral detections in general are kind of table stakes in a lot of tools.”

CISO, business advisory and financial services organization

The absence of a known-bad signal is not the absence of risk. Asked which approach primarily detects suspicious activity in that situation, respondents answered:

Behavioral anomaly detection was chosen nearly as often as the next two put together. Known malicious indicators still matter, but behavior asks a second question: is it normal?

Most organizations have already embraced behavioral baselining. The harder problem is establishing a baseline of normal behavior precise enough to distinguish meaningful deviation from the ordinary variability of enterprise activity.

Yet the two most commonly cited barriers to expanding it are:

“There's not a consistent flow of behavior when it comes to humans, that we can say, this is what normal looks like all the time.”

CISO, healthcare organization

For that organization, the problem is concrete. Its identity population includes employees as well as physicians and other authorized users whose activity can vary dramatically.

That variability is what turns baselining from a deployment problem into a context problem. The more normal activity changes by person, role, application, location, or circumstance, the harder it becomes to draw a useful boundary between unusual and risky.

“Too many false positives start to erode confidence.”

CISO, healthcare organization

There Is No Single Formula for Normal

Automatically learned baselines are common in mature environments, while predefined rules continue to play a substantial role alongside them. Organizations frequently use both rather than treating behavioral learning as a replacement for established logic.

Some behaviors can be bounded with explicit policies; others require systems to learn what is typical over time. Either way, the test is whether a deviation is merely unusual or actually suspicious.

That distinction becomes especially important when an attacker does not need to break through a control at all.

“You're not being hacked. They're logging in.”

CISO, healthcare organization

The same CISO described attackers using legitimate credentials and then appearing normal once inside the environment.

Detecting deviation is only useful if the baseline contains enough identity and environmental context to tell the difference.

Finding an anomaly is only the first step. The harder operational test is whether security teams can understand what happened, determine whether it represents real risk, and respond quickly enough to matter.

On the surface, trust is relatively high: 65% say analysts trust behavior-based detections enough to act without additional manual validation always or most of the time. But unconditional trust is rare—only 8% say analysts can always act without it.

The Understand-to-Act Gap

1. Detect
2. →Understand
3. →Trust
4. →Act
5. →Resolve

“All we know is it's misbehaving...we have privacy officers, we have HR, we have clinicians look at that and decide if that's normal behavior or not, and then take action.”

CISO, healthcare organization

Investigation Still Takes Time

That time reflects more than technical investigation. Behavioral detections frequently require context: who or what generated the activity, what that identity normally does, whether permissions or responsibilities recently changed, and whether an unusual action has a legitimate explanation.

“This is how we learn… by sharing information… But that's… a very manual, serial process.”

CISO, healthcare organization

At enterprise scale that model becomes difficult to sustain. Another CISO described the volume of security data as "phenomenal," arguing it can no longer be handled without some degree of automation and AI-assisted detection. Organizations appear to recognize the bottleneck.

Visibility Is Not Understanding

79% of respondents say they can easily produce a unified view of anomalous activity across all major communication and identity environments.

But aggregating anomalies is not the same as explaining them.

Behavioral security may be mainstream, but its reach is uneven.

That difference matters because non-human activity is not peripheral to modern organizations. Applications, workloads, APIs, service accounts, and other machine identities routinely access systems and data without a person directly performing each action.

“The non-human identities… are really where the new playground of risk is surfacing.”

CISO, B2B software organization

The Machine Identity Confidence Paradox

The coverage gap does not translate into lower perceived preparedness. 72% say their organization is somewhat or much better prepared to detect anomalous behavior from non-human identities than from human identities—yet coverage remains incomplete, and effectiveness is uneven.

The findings do not necessarily mean security teams are mistaken about their preparedness. One possible explanation is that many traditional non-human identities still perform narrower, more predictable functions than people do. A service account with a defined purpose and set of permissions may make deviation easier to recognize even if monitoring around it is less comprehensive. AI agents are bound to take on far more advanced functions.

“We know what a service account's supposed to do and what it's not supposed to do. The challenge with [AI] agents is: we don't know what we don't know.”

CISO, healthcare organization

Security teams are making their priorities clear. Asked to identify the single area that will receive the greatest behavioral-monitoring investment over the next 24 months, 43% chose AI agents. The next-highest category—API integrations—received just 14%. That is not a narrow lead.

The market is not starting from zero. 54% already maintain behavioral baselines around AI agents, and 31% name AI-agent coverage as a top-three priority for improving behavioral baselining.

The AI Agent Investment Paradox

Only 42% rate themselves very or extremely effective. The tension is not simply that AI-agent security is immature—it is that organizations are committing investment before they have worked out what effective monitoring should look like.

“It is the number one planned investment for practically every CISO I've talked to… for next year. There is no plan...we don't know what the solutions are, but we're all gonna put resources to go after it.”

CISO, B2B software organization

Autonomy Changes the Behavioral Question

The challenge is not simply extending existing machine-identity controls to one more category. AI agents can act on behalf of people, access information across multiple systems, make decisions about how to complete assigned tasks, and generate additional actions that were not individually specified by a human. One CISO described agents already beginning to blur conventional attribution:

“It's blurring the lines between what is truly the agent and what is the human, because it's acting on behalf of people.”

CISO, B2B software organization

That complicates a security model built around relatively stable relationships between identity, access, and action. The security problem therefore becomes more nuanced than determining whether an agent is authorized to exist. It becomes:

When the Actor Itself Is Dynamic

Agent behavior can also be temporary and difficult to attribute after the fact. One CISO distinguished relatively durable agents—which can be tied to a user or service principal—from ephemeral sub-agents created for a single task that exist only briefly.

“How are you gonna go audit an AI activity that lasted for 30 seconds on a virtual machine?”

CISO, business advisory and financial services organization

That organization could still observe underlying network activity. But the CISO drew a distinction between seeing traffic and being able to attribute it specifically to an agent—a problem that becomes harder when the actor is short-lived or buried several layers deep in another process.