54% of enterprises hit by AI agent security incidents; most still share credentials

VentureBeat Pulse Research surveyed 107 enterprises to examine how organizations are securing their AI agents, looking at the tooling they deploy, identity and isolation management practices, and incidents that have already occurred.

The findings paint a picture of widespread exposure. Across the surveyed sample, 54% have already had a confirmed AI agent security incident or experienced a near-miss—a striking figure that indicates the problem is already affecting a majority of organizations attempting to deploy agents in production. This occurs even as enterprises grant agents real access to systems and data, making containment essential.

Yet the containment measures remain underdeveloped. Only about a third of enterprises give every agent its own scoped identity; the majority still allow agents to share credentials, a practice that complicates auditing, revocation, and forensic analysis. Even more concerning, only three in ten isolate their highest-risk agents—the very systems where the damage from a compromise or misconfiguration would be greatest.

The tooling picture reflects this immaturity. The security stack deployed across these enterprises is overwhelmingly borrowed from model providers and hyperscalers—companies like AWS or Azure or the makers of large language models—rather than purpose-built specifically for agent deployment patterns. This approach, while expedient, leaves gaps tailored to agent-specific threats. Furthermore, spending on agent security remains a thin slice of the overall security budget, suggesting that organizations have not yet prioritized dedicated investment in this area.

On defense posture, enterprises themselves are uncertain. The survey found them evenly split on whether their current defenses are keeping pace with AI-enabled attackers. As the body notes, autonomous agents are proliferating faster than the identity, isolation, and enforcement controls needed to contain them—a gap described as the "agent security gap." The research ultimately points to a pressing need for controls that match the speed and scope of agent deployment.

The research surface a critical mismatch in enterprise AI deployment. Organizations are granting AI agents real access to systems and data—a necessary step for productive use—but the security infrastructure to constrain that access has not kept pace. The fact that 54% of enterprises across a 107-company sample have already experienced an incident or near-miss suggests the gap is not theoretical; it is already manifesting in real operational risk.

Several structural factors contribute to this gap. First, identity and isolation practices remain immature: only about a third of enterprises have assigned each agent a scoped identity, and most still rely on shared credentials—a practice that violates least-privilege principles and makes forensics and revocation much harder. Second, the tooling landscape is dominated by solutions adapted from model providers and hyperscalers, not purpose-built for agent deployment patterns. Third, and crucially, agent security spending remains marginal within the overall security budget, suggesting that the problem is not yet perceived as urgent enough to attract dedicated investment.

The fact that enterprises are evenly split on whether defenses are keeping pace with attackers underscores the uncertainty. As autonomous agents continue to proliferate—and the body states they are proliferating faster than controls can scale—the risk surface will only expand. The research points to a window in which purpose-built tooling, clearer identity standards, and intentional isolation architecture could help close the gap before the problem becomes endemic.

No comments yet. Be the first to share your thoughts!