Cloud Security Cloud Security Posture Management

Security teams lack baselines for AI agent behavior, report finds

Many organizations are deploying AI agents without a reliable picture of how those systems normally behave, making it harder for security teams to identify activity that may be unsafe or malicious. In a report published last week, security firm Abnormal AI found that 46% of surveyed security professionals said they could not establish safe baselines for agentic AI activity.

The findings point to a governance problem that is growing alongside enterprise use of autonomous tools: monitoring systems need a reference for expected behavior, but organizations have not yet established one consistently. Without that reference, teams may struggle to distinguish legitimate agent actions from unusual activity, including actions that could indicate misuse, compromise or data theft.

Machine identities add to the challenge

More than half of respondents, 53%, said their organizations did not establish normal behavior baselines for machine identities, a category that includes AI agents. At the same time, interest in monitoring these systems is increasing. More than four in 10 businesses investing in behavioral-monitoring technology were directing it toward AI agents, according to Abnormal AI. Nearly one in three respondents identified agentic AI as one of the most important areas in which to improve baselining.

That combination—high interest and limited readiness—suggests that planned investment does not necessarily translate into an operational approach. A security executive at an unnamed B2B software company told the report’s survey-takers that behavioral monitoring was a leading planned investment for security leaders, but that organizations lacked a clear plan for making the investment effective. The executive’s comments reflect uncertainty about which solutions will work, even as companies intend to commit resources.

AI agents complicate established security models because they can act both on behalf of people and on their own initiative. Abnormal AI said this blurs the distinction between human and machine behavior, challenging models built around relatively stable links among identity, access and action. An agent may have permission to perform an operation, yet permission alone does not establish whether that operation is appropriate in the context of the agent’s instructions.

That distinction matters for security monitoring. Organizations need to assess not only whether an identity has authorization, but also whether its activity fits its assigned role and expected behavior. A baseline can help provide that context, though the report’s findings indicate that many companies have not yet established one for agents or other machine identities.

Alert confidence remains a barrier

Defining normal behavior is not the only obstacle. Nearly four in 10 respondents said difficulty understanding or trusting behavioral alerts was a key challenge in monitoring network activity. Abnormal AI identified three leading hurdles: uncertainty about what normal behavior looks like, lack of trust in alerts, and an overwhelming volume of false positives. The firm said these issues outranked more conventional constraints such as staffing, budgets and tool capabilities.

When alerts are difficult to interpret or frequently prove false, security teams may have less confidence in the monitoring process. In the case of AI agents, that can compound the baseline problem: teams need reliable signals to learn what activity is expected, but questionable alerts can make it harder to determine which deviations merit investigation. The report describes these as connected challenges for organizations seeking to analyze behavior across their networks.

Abnormal AI interviewed 525 security professionals in six countries for the report. Each respondent represented an organization with at least 1,000 employees, and 85% were chief information security officers or security operations center leaders. The results therefore capture views from security decision-makers and operations leaders at larger organizations, rather than a measure of how all businesses are managing AI agents.

The report does not prescribe a single technical fix. Its findings instead underscore the work required to make behavioral monitoring useful: establish expectations for machine identities, assess actions against an agent’s instructions, and improve confidence in alerts. As organizations expand their use of agentic AI, those steps will be central to determining whether security teams can recognize behavior that falls outside an agent’s intended role.