AI Agent Security Incidents: Over Half of Firms Hit Already

Here’s the shocker: More than half of surveyed enterprises have suffered a confirmed security incident or a close call involving AI agents. Yet, most still allow these agents to share credentials. For IT and security leaders, this isn’t just a minor oversight—it’s an open invitation to attackers. Shared access means a single compromised agent can jeopardize multiple systems in one go. The business risk isn’t theoretical: we’re seeing real-world cases already. The problem is less about futuristic threats and more about basic hygiene gone missing amid the AI rush.

Only about a third of organizations assign each AI agent a unique, scoped identity. In practice, this means that when something goes wrong, it’s difficult to tell which agent did what—or even to cut off a single compromised agent without disrupting others. Even among companies that use agents for sensitive or high-risk tasks, only three in ten bother to isolate those agents. The result? Attackers can move laterally through agent networks with relative ease. This gap isn’t just theoretical. In the projects we run, we've noticed how businesses underestimate the consequences of poorly defined agent boundaries until faced with a costly forensic investigation.

Most enterprises rely on whatever security features are bundled by their AI model providers or hyperscalers. Out-of-the-box controls might suffice for basic application hosting, but AI agents have unique behaviors and risks. Purpose-built security—controls designed specifically for autonomous agents—remains rare. This leaves glaring blind spots around agent autonomy, context switching, and delegated authority. Enterprises are treating agent security as an afterthought, not a budget line. Until that changes, expect incidents to keep piling up.

Spending on agent-specific security is barely a sliver of the overall security budget. This tells us that organizations still see AI agents as experimental, not mission-critical infrastructure. Yet the operational reality is shifting: agents are increasingly wired into core business systems and data. Leadership’s confidence is evenly split: half believe their defenses can keep up with AI-powered attacks, the other half clearly don’t. Such ambivalence breeds risk. Without dedicated investment, even robust general-purpose defenses can’t anticipate the unique exploits that autonomous agents attract.

The pattern is clear: agent autonomy is outpacing basic controls. Businesses must stop treating AI agents like traditional scripts or service accounts. Every agent needs its own identity, strict isolation, and purpose-built enforcement layers—especially for those with access to sensitive data or critical operations. Relying on default cloud tooling will buy time, but not immunity. As incidents mount, the cost of retrofitting security after the fact will dwarf any upfront investment. Enterprises that move now to close this gap will be better positioned to unlock real business value from AI without courting disaster.