AI Agents Proliferate, Outpacing Governance Controls

Fifty-eight percent of organizations operate more than 50 AI agents today, according to a survey of 251 security leaders reported by Zentera Systems.

Rapid growth of AI agent fleets in surveyed organizations

The Zentera Systems survey — 251 security leaders across multiple sectors — paints a fast-moving picture. Today, 58% of respondents say their organization runs more than 50 AI agents. That number is expected to climb: within 12 months, 66% anticipate operating over 50 agents and 38% expect to exceed 100. Deployment is not solely grassroots; 36% of leaders strongly agree that company leadership has directed the deployment of AI agents or agentic capabilities.

Authorization as the missing layer in agentic AI security

Respondents identify clear governance shortfalls. Eighty-seven percent agree that authorization is the missing layer in agentic AI security. Leaders worry not only that agents will act, but that they will act where and how they should not: 75% are concerned an AI agent could perform the correct task in the wrong environment or location, and 80% are concerned agents may hold access that was never explicitly granted. The survey highlights a specific tension: "A technically correct action can still be unauthorized because of where it happens and what it reaches."

Desired controls: explicit authorization, project isolation, session logging

When asked which controls they want before expanding agent fleets, leaders rank explicit authorization first (56%), followed by project isolation (51%) and session logging (48%). Support for isolation is strong across the sample: 85% say project-level isolation is essential or very essential for AI adoption. Those priorities indicate that security teams see authorization and compartmentalization as the immediate levers to reduce the risk of agents “crossing boundaries.”

Auditability and monitoring fall short of confidence

The survey suggests many organizations lack high-confidence oversight even as agents proliferate. Only 43% of leaders say they are very confident they can demonstrate what AI agents were explicitly authorized to do, and 38% are very confident they can prove what an agent did through audit records. Just 37% monitor agent activity very closely. Zentera’s summary notes that for each form of oversight measured, most leaders “stop short of the highest confidence tier,” signaling gaps between desired controls and current practice.

What this means for technologists, procurement leaders, and regulators

• Technologists and security teams: they must prioritize the technical controls leaders ranked highest — explicit authorization, project-level isolation, and session logging — because the survey shows those controls are viewed as essential to safe expansion of agent fleets.
• Procurement leaders and business executives: the data shows active leadership direction to deploy agents (36% strongly agree), but the same leaders are concerned about unintended access and boundary-crossing; procurement choices will need to reflect the demand for authorization and isolation features.
• Regulators and governance bodies: expectations of rollback are high. Seventy-nine percent of respondents expect their organization will need to claw back or significantly restrict AI agent usage within the next 18 months, and Gartner separately predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because governance gaps are identified after production incidents occur.

The narrow takeaway is clear: many organizations already run large fleets of agents, and most security leaders believe the systems that control where and how those agents operate lag behind. The survey’s figures — acceptance of project isolation as essential, a broad call for explicit authorization, and the lack of high-confidence auditability — together make a simple prediction explicit in the reporting itself: expansive fleets without tighter authorization and isolation will likely prompt contraction. Whether companies will act before Gartner’s forecasted demotions and decommissions materialize depends on whether the authorization layer moves from “missing” to mandatory.

Read the full report at https://www.securitymagazine.com/articles/102584-majority-of-organizations-have-over-50-ai-agents.