Security leaders now view autonomous AI agents with unchecked access as the single greatest danger to their organizations, according to a survey commissioned by Exabeam and conducted by Sapio Research. The finding places a new class of non-human actor at the top of the threat list, ahead of the external attackers and insider risks that have traditionally dominated enterprise risk registers.
The numbers behind the shift
Among the 600 security and finance decision-makers polled, 48% identified autonomous AI agents with excessive access as their foremost concern. External hackers ranked second, chosen by 28% of respondents. Compromised insiders and malicious employees each drew 12%, placing them well behind both the agentic threat and the external attacker category.
The survey drew from organizations with 500 or more employees and spanned the United States, Canada, the United Kingdom, France, Germany, the Netherlands, and Australia. The results reflect perceived risk rather than confirmed incident data, a distinction that matters when interpreting the ranking.
What makes an AI agent different
Exabeam defines AI agents as goal-driven autonomous systems that operate with limited human oversight. Unlike conversational chatbots, which respond to prompts and generally stay within a narrow interaction boundary, agents are designed to pursue objectives across systems. To do that, they rely on legitimate administrative credentials, API keys, and elevated privileges.
That access profile is precisely what makes them difficult to police. Because agents authenticate with valid credentials, their actions can bypass perimeter defenses and slip past behavioral baselines that were built to flag anomalous human activity. A rogue or misconfigured agent may not trigger standard security alerts at all.
Critically, the risk does not require malicious intent. An agent with unchecked access can cause harm through error, drift, or misconfiguration alone. The exposure is structural: it lives in the gap between granting access and overseeing how that access is used.
Monitoring gaps and the analytics response
Exabeam’s report, titled The Agentic Insider: From Monitoring to Understanding, addresses the difficulty organizations face as they expand AI agent monitoring. Companies are adding oversight capabilities, but many still struggle to understand what agents actually do across systems. Tracking activity is one problem; interpreting it is another.
Security vendors are responding by building agent behavior analytics, tools intended to establish what normal agent activity looks like and to surface deviations. The survey results suggest that demand for such capabilities is likely to grow as autonomous systems take on more responsibility inside enterprises.
Perceived risk, not confirmed incidents
The survey measures how decision-makers perceive the threat landscape, not how many breaches have been traced to AI agents. That caveat is important. A high ranking for agentic risk signals a shift in priorities among security leaders, but it does not by itself prove that agents are causing incidents at a rate matching the concern they generate.
Still, the ranking is notable for what it displaces. External hackers, long the default top concern, now sit a distant second. Insider categories that have driven years of investment in monitoring and access controls rank lower still. The ordering suggests that security teams are rethinking which actors deserve the most attention.
Why the timing matters
The survey arrives as organizations expand their use of autonomous AI and as ransomware numbers continue to rise. IT teams are already stretched managing Windows updates and patches, including the recent Microsoft preview update KB5124010, which introduced new features alongside two known issues. Adding AI agent oversight to that workload increases the pressure on administrators.
The practical implication is that oversight must extend beyond operating systems and traditional endpoints to the agents themselves. Administrators who already track patch status and system behavior may need to apply similar discipline to the credentials, permissions, and actions of autonomous systems operating inside their environments.
What security teams should take away
The Exabeam survey points to a threat category that does not fit neatly into existing frameworks. AI agents are not outsiders, nor are they employees. They are authorized actors whose access is legitimate and whose behavior can be opaque. That combination is what makes them hard to detect and hard to govern.
For security leaders, the message is less about any single percentage than about where attention is moving. If nearly half of surveyed decision-makers already rank agentic access as their top concern, the operational question becomes whether monitoring, analytics, and access governance have kept pace with the autonomy being deployed.
