Most Organizations Running AI Agents: Security Gaps and Risks Exposed by New Report

| Editorial Team On 22 Oct, 2026

Most Organizations Running AI Agents Lack the Security to Protect Them

A new ServiceNow report reveals that 90% of global organizations have suffered an agentic AI security incident in the past year — yet only 13% have defensible security programs in place to handle the threat.

The findings expose a widening gap between how secure organizations believe they are and how secure they actually are. As AI agents multiply across enterprise environments at machine speed, human-driven governance frameworks are struggling to keep pace — leaving businesses exposed to risks that move faster than their ability to respond.

The Scale of AI Agent Deployment Is Outrunning Security Readiness

The numbers alone tell a sobering story. According to the ServiceNow report released October 21, 2026, 49% of organizations now run between 100 and 999 AI agents simultaneously. A further 35% operate more than 1,000 AI agents within their environments.

That level of deployment demands an equally scaled security response. Yet the infrastructure to support it is largely missing.

To understand the breadth of how AI is now embedded in business operations, it helps to look at how businesses are actively deploying artificial intelligence across different functions and sectors — the proliferation is significant, and the security implications grow alongside it.

Just 44% of security teams globally design security into an agent's development before it is built. For the majority of organizations, that means permissions and guardrails are applied after the fact — reactive patches on systems already in motion.

Most agentic AI today is governed by legacy frameworks built for slower, more human-driven systems. Less than half of organizations use governance frameworks that cover the key elements of AI agent deployments. The result is a sprawling AI footprint with outdated guardrails trying to contain it.

The Confidence Gap Is Wider Than Most Leaders Realise

When measured against five key domains — attack surface, accountability, security by design, visibility, and audit readiness — only 13% of organizations have what the report considers a truly defensible agentic security program. That figure stands in sharp contrast to the far higher number of security leaders who self-assess as having established security postures.

Confidence is not the same as capability.

This distinction matters enormously. An organization that believes it is protected but lacks the underlying frameworks to validate that belief is, in practical terms, more exposed than one that openly acknowledges its gaps and works to close them. Self-assessed security posture has long been a problematic metric, and agentic AI is making the consequences of that problem significantly more serious.

One of the report's most striking findings involves what organizations simply cannot see. While 57% of security teams say they maintain a full and accurate record of every AI agent in production, 29% of teams have discovered AI agents running in their environments that were never registered or officially known.

These are shadow AI systems — agents operating unsupervised, outside sanctioned processes, and invisible to the security teams responsible for protecting the organization. In the same way shadow IT created compliance and security headaches a decade ago, shadow AI is emerging as a parallel threat that is arguably more dangerous given the autonomous nature of AI agents.

What gets missed tends to surface only after damage is already possible. An unregistered agent with broad permissions and no oversight is not a theoretical risk. It is an open door.

Excessive Permissions Are Amplifying the Problem

The problem compounds when organizations consider how AI agents interact with sensitive systems and data. The report notes that only 33% of AI agents are provisioned with least-privilege access — a foundational cybersecurity principle that limits what any given system or user can access to only what is strictly necessary.

When AI agents are granted excessive permissions and then run without visibility, the exposure multiplies quickly. The risks and challenges that artificial intelligence presents to business operations are well-documented, but the agentic dimension adds a layer of urgency that static AI deployments do not carry. An agent that can act, decide, and execute autonomously — with excessive access — represents a category of risk that most current governance models were never designed to address.

Machine-Speed Risk Is Outpacing Human-Speed Response

Perhaps the most urgent finding in the report is the speed mismatch between how fast AI agents can cause harm and how fast humans can intervene.

76% of organizations report maintaining a complete and current map of every agent's access. However, only 19% can see more than 90% of their agents in real time. Visibility on paper is not the same as operational awareness in the moment.

The Response Gap Remains a Critical Vulnerability

The response gap is even more concerning. Two-thirds of organizations — 66% — cannot act automatically when an agent behaves anomalously. That means a human must still detect a problem and manually intervene before any corrective action can take place. In environments where AI agents execute tasks at machine speed, that human bottleneck can turn a containable incident into a significant breach.

Of the organizations that did experience an agentic AI security incident in the past year, 36% were able to contain it within 24 hours. That leaves nearly two-thirds taking longer — or failing to contain it entirely within that window.

The analogy here echoes a familiar challenge from cybersecurity history. Just as organizations once struggled to respond to automated malware spreading faster than human analysts could track, the agentic AI era is introducing a new version of that race — one where the autonomous systems in question may be operating with legitimate credentials and sanctioned access. That distinction makes detection harder, not easier.

For organizations actively pursuing AI-led business transformation strategies, the security implications of agentic deployment cannot be treated as secondary considerations. The faster transformation moves, the faster the attack surface expands — and the more critical it becomes to have security infrastructure that moves at the same pace.

The NIST AI Risk Management Framework provides one of the most comprehensive publicly available resources for organizations seeking structured guidance on managing AI-related risk — and is worth incorporating into any agentic security programme.

What This Means for Your Organisation

The ServiceNow report offers a clear picture of where the enterprise AI security landscape stands heading into late 2026. For security leaders and technology decision-makers, the findings carry direct implications.

Audit your AI agent inventory now. If 29% of security teams are discovering unregistered agents in production, the first step is knowing exactly what is running in your environment. An accurate and continuously updated agent registry is not optional — it is foundational.
Shift security left in AI development. With only 44% of teams building security in before deployment, the majority of organizations are creating risk before their agents ever go live. Integrating security by design into agent development pipelines reduces the need for reactive fixes later.
Invest in automated response capabilities. Human-speed intervention cannot match machine-speed risk. Organizations that cannot automatically respond to anomalous agent behaviour are structurally exposed. Building automated detection and response workflows for agentic AI should be a near-term priority for security teams navigating this landscape.