AI agents topped the list of planned investments in behavioral monitoring in Abnormal AI’s 2026 State of Behavioral Security Report. Asked where their organizations would invest most over the next two years, 43% of respondents chose AI agents, more than three times the share who chose API integrations, the next most common answer.
Security professionals were divided on how well their organizations could establish and maintain baselines for normal AI agent behavior. Nearly half said their organizations were not at all effective, while about two in five rated them very or extremely effective.
Monitoring extends to non-human identities
More than 81% of respondents said their organizations rely on behavioral anomaly detection to a major or extensive degree. The approach uses patterns of activity to flag events that may warrant investigation. A sign-in outside someone’s usual pattern, for example, can draw attention even when the credentials are valid.
The same approach can apply to non-human identities, which software uses to authenticate and access systems. A service account might run a scheduled backup, an API integration might exchange data between applications, and an agent might retrieve information on someone’s behalf. Each has a different pattern of expected activity.
Behavioral baseline monitoring coverage across systems and activities (Source: Abnormal AI)
About 72% of respondents said their organizations were better prepared to detect unusual behavior from non-human identities than from people. In a separate question, 76.2% reported maintaining baselines for user sign-ins and authentication. Coverage was lower for AI agents, API integrations and service accounts, with each category at roughly half.
Security teams use past activity to help identify actions that look out of place. That can be difficult even for people with familiar roles. One healthcare security leader said an authorized user might access the organization’s systems several times in three months, then go six months or a year without returning. A system that treats a return after a long absence as suspicious can flag someone whose access is part of their job.
What counts as normal depends on the task
An agent can act on someone’s behalf, gather information from several systems and choose steps the person did not specify. It may access a system that the person who assigned the task does not use directly. To assess that activity, a security team may need to know what the agent was asked to do, whose permissions it used and how the action fit the assignment.
One software company’s security chief used the example of AI activity lasting just 30 seconds on a virtual machine to explain how hard it can be to audit a short-lived agent. A security team may be able to see the network traffic and still struggle to tell which agent generated it, especially if the agent was created for a single job or operated several layers deep in another process.
Investigating unusual activity can take hours
Just over half of respondents said their organizations typically needed at least four hours from detecting a behavioral anomaly to completing the investigation and remediation. That timing covers behavioral anomalies in general, including activity unrelated to AI agents.
Respondents most often cited false positives and difficulty defining normal activity as obstacles to expanding behavioral baselining.
An alert can tell a security team that something changed. Investigating it requires enough information to explain the change. A record showing that an agent opened a file captures the action. The assignment helps explain whether the agent had a reason to open it.
