AI Agents Are in Production. Confidence Isn't.
New Omdia research on how 200 engineering and security leaders are deploying AI agents—and why containment is what provides confidence around their security.
eBook
Agents Are Scaling Faster Than the Controls Around Them
AI agents are already everywhere: 96% of organizations have them deployed, and they expect their production agent count to grow 46% on average over the next 12 months. Controls, however, haven't kept pace. Among teams that experienced an agent failure, 34% saw agents execute unauthorized actions, yet only 23% of organizations use hardware-enforced isolation today.
In this eBook, Omdia surveys 200 engineering and security leaders on where agents are running, what's breaking, and what it takes to deploy them with confidence.
eBook
Where Agentic AI Breaks & Where It Holds
Teams are shipping faster than ever and AI agents are adding to both the output and the security load. As agent deployments grow just a couple of agents to 100+, developers are spending more of their time on security. This Omdia eBook maps where agent deployments break down and why 89% agree that containment reduces the risk you can't patch your way out of.
Inside:
- Where organizations are putting agents to work, from security operations to code generation.
- The top blockers to agent deployment.
- What goes wrong when agents fail.
- How organizations isolate untrusted workloads today.
- Why security friction grows with your agent fleet, and what it takes to scale adoption without scaling risk.
Frequently Asked Questions
Adoption is nearly universal. 96% of surveyed organizations have AI agents deployed, with 46% average expected growth in production agents over the next 12 months. The most common use cases are security operations (66%), IT infrastructure management (59%), data analysis (50%), and code generation (48%).
Data privacy, managing access controls and permissions for autonomous agents, and integration complexity top the list. More than half of respondents rate data privacy a major obstacle or critical blocker, and say the same about preventing unintended interactions between agents or workloads.
Among respondents who experienced an agent failure, 34% saw agents execute unauthorized actions or commands, 27% cited insufficient isolation for agent workloads, and 24% saw blast radius expansion, where agents reached resources beyond their intended scope. Another 24% found they needed hardware-enforced isolation beyond traditional boundaries.
Prevention can't catch every failure, especially when non-deterministic agents run code at machine speed. Containment limits how far a compromise can reach. Most respondents agree containment practices reduce risk that can't be "out-patched," and many say at least 31% of their production workloads are untrusted or need enhanced isolation. Yet most still rely on container-level isolation or network-level controls alone; only 23% use hardware-enforced isolation.
Omdia conducted the survey on Edera's behalf, polling 200 US professionals—140 engineering leaders and 60 security and CISO-level professionals—at organizations with more than 1,000 employees. The survey ran Aug 4–14, 2026, with a margin of error of ±6.5%.
Additional Resources
A Curated Collection of Musings & Research

Let's Build Together
Building AI Agents?
Let's explore how Edera fits into your Kubernetes environment and the agent platform you're building.



