AvePoint and Osterman Research Surveyed 750 IT Leaders. 88% Had an AI Agent Security Incident. Most Still Call Themselves Prepared.

October 22, 2026 · Josh Agusti, Founder, Aiven Consulting Group

Most executives will tell you their organization is on top of AI security. A new survey suggests that confidence and reality have quietly come apart. AvePoint, working with Osterman Research, surveyed 750 IT and security leaders across the Americas, EMEA, and APAC for its State of AI 2026 report, and the headline finding is stark.

About 88 percent of respondents said their organization had experienced at least one AI agent related security incident in the past twelve months. That is not a fringe problem affecting a handful of early adopters. It is close to the norm.

The confidence that doesn't hold

The more uncomfortable finding sits underneath that headline number. Over 80 percent of respondents told AvePoint they were confident in their organization's ability to prevent unauthorized access involving AI systems. But among that confident group, somewhere between 62 and 72 percent still reported unauthorized access incidents tied to AI. Confidence, in other words, was not a reliable predictor of safety. Leaders who believed their controls were working were nearly as likely to have been breached as leaders who had no such belief. That is the pattern worth sitting with. A governance program that produces confidence without producing evidence is not actually reducing risk, it is just changing how comfortable people feel while the risk stays the same.

Why agents create a different kind of blind spot

Part of what makes this harder than traditional security governance is that AI agents do not fail the way old systems failed. A misconfigured server or an unpatched application tends to leave a visible trail. An agent that was spun up by a business team to automate an email workflow, and that now reaches into a shared drive or a CRM on its own, often leaves no trail at all until something goes wrong. AvePoint found that roughly one in five respondents could not say whether staff in their organization were creating AI agents without any formal approval process. A separate 2026 report from Pathlock, surveying 286 IT, compliance, and security leaders, found a similar pattern from a different angle: more than half of respondents were not confident they knew about every AI agent operating inside their business systems. Two independent surveys, two different populations, landing on the same basic conclusion. Most organizations evaluating agentic AI use cases are moving faster at adoption than they are at keeping an inventory of what they have actually deployed.

From confidence to evidence

The fix is not a new confidence survey. It is a different kind of proof. Mature AI governance programs are starting to look less like a policy document that gets signed once a year and more like an operating discipline: every agent has a named owner, every agent's access is scoped to what its task actually requires, every agent's actions are logged somewhere a human can review them, and incident response plans explicitly cover what happens when an agent, not a person, is the one that made the mistake. None of that is exotic. It is the same discipline security teams already apply to service accounts and automated scripts, extended to a new category of actor that happens to make its own decisions. Organizations that have put a governance framework in place before scaling agents tend to discover problems through their own monitoring rather than through an incident.

What this means for leadership right now

The practical takeaway for an executive team is not to slow down AI adoption. It is to stop treating confidence as a substitute for an inventory. Ask a simple question in the next leadership meeting: can anyone in the room name every AI agent currently running in a core business system, who owns it, and what it is allowed to touch. If the honest answer is no, that is the actual governance gap, regardless of how prepared anyone feels. The organizations AvePoint surveyed that had the fewest incidents were not the ones with the most confidence. They were the ones that had already done the unglamorous work of finding out what they did not know.

Key Takeaways

AvePoint and Osterman Research's State of AI 2026 report, based on a survey of 750 global IT and security leaders, found that about 88 percent of organizations had at least one AI agent related security incident in the past 12 months, even though most leaders describe themselves as confident in their AI security posture.
Confidence and control turned out not to be the same thing: among organizations that said they were confident in their ability to prevent unauthorized access, a majority still reported unauthorized access incidents tied to AI systems.
A meaningful share of leaders admitted they could not say whether every AI agent running inside their organization had been formally approved, which means governance has to shift from periodic attestation to continuous, evidence based oversight.