The biggest AI risk for CIOs may not be the model itself. It is the systems forming around it faster than IT can govern them.
AI agents are beginning to act inside enterprise environments, calling APIs, using credentials, accessing data and executing workflows. In many cases, they’re being deployed by business units and application teams before security, IT and governance processes have caught up. That creates a familiar problem in a new form. Enterprises are building a layer of automated activity that looks legitimate on the surface, but may be difficult to inventory, monitor or control.
Radware’s 2026 Cyber Survey, conducted by Osterman Research, shows how wide that gap has become. More than 83% of organizations report widespread use of GenAI or LLM functionality, and 96% have implemented or plan to implement AI agents or autonomous workflows within the next 12 months. Yet only 17.2% report complete visibility into the AI agents and AI-driven processes operating in their environments.
For CIOs, this is no longer a future-planning issue. AI is already becoming part of how applications run, how employees work and how business processes move. The question is whether the organization can still see, govern and secure the activity it is creating.
AI Agents Create an Identity and Authorization Problem
As agents move into production, they become more than software features. They become machine identities with permissions.
An agent may authenticate to an application, call several APIs, retrieve sensitive data, invoke other tools or initiate a business transaction. Depending on how it is configured, it may perform those actions without a person approving each individual request.
Without an accurate inventory, security teams cannot reliably answer which agents are running, whose credentials they use, which APIs they can reach, what information they can access or what actions they are authorized to perform.
The risk comes from the gap between deployment and control. An agent can behave exactly as designed and still create exposure if its permissions are excessive, its credentials are compromised or its access extends into systems that were never considered when it was first deployed. For CIOs, AI governance has to extend past approving models or use cases and needs to include identity, access, permissions and runtime behavior.
Also Read: CIO Influence Interview with John Elliott, Cybersecurity Author Fellow at Pluralsight
Automated Traffic Is Getting Harder to Classify
A similar problem is appearing at the application edge. Only 14% of organizations surveyed report complete visibility into AI crawler traffic accessing their applications. Just 18% say they are extremely confident in their ability to distinguish legitimate AI agents from malicious bots.
Automated traffic can’t only be classified as bot or human. An incoming request may come from a legitimate search crawler, an AI training crawler, a retrieval service, an autonomous agent acting on behalf of a user or malicious automation pretending to be one of them. Those requests may look similar at the HTTP layer while having very different purposes and access requirements.
And it makes enforcement harder. Blocking too aggressively can interfere with legitimate services. Allowing traffic based only on a claimed identity can expose applications, proprietary content and APIs to scraping, reconnaissance or abuse. Defenders need to classify automated activity based on identity, behavior and the resources being accessed, rather than relying on a binary human-versus-bot decision.
API Change Is Outrunning API Discovery
Nearly half of organizations now update production APIs daily or multiple times per day, but security practices are lagging. Only 19% of organizations maintain a fully automated, continuously updated inventory of their APIs. Just 24% conduct comprehensive API security testing throughout the development and production lifecycle, and fewer than one-third continuously monitor API posture and configuration risks across most or all APIs.
When production APIs change faster than discovery, security teams are working from an outdated view of the attack surface. Organizations can’t effectively secure APIs they haven’t discovered, documented or continuously monitored.
Control Has to Catch Up With Deployment
Treat AI agents as production actors, not just productivity tools. Organizations need to know which agents exist, what identities and credentials they use, which APIs and data they can access, and what actions they are allowed to perform. They also need continuous API discovery, better classification of automated traffic and security operations capable of correlating and responding to automated activity quickly.
AI adoption is unlikely to slow while enterprises solve these problems. If an organization can deploy an agent faster than it can identify that agent, understand its permissions and monitor its behavior, AI adoption is already moving faster than the controls designed to manage it.
About The Author Of This Article
Dan Schnour is Senior Product Manager, Radware
About Radware
Radware® is a global leader in application security and delivery solutions for multi-cloud environments. The company’s cloud application, infrastructure, API, and AI security solutions use AI-driven algorithms for precise, behavior-based, real-time protection against sophisticated web, application, and DDoS attacks, API abuse, business logic threats, and malicious bots. Radware delivers end-to-end API security, including discovery, posture management, testing, and runtime protection, along with advanced protection for AI agents and models.
Catch more CIO Insights: How Are CIOs Aligning Technology with Workforce Agility?
[To share your insights with us, please write to psen@itechseries.com ]
