How confident are you that you know about every AI agent running in your environment right now? 

If you’re like most enterprise security leaders, the honest answer is “not very.” You probably have a handle on the official copilots and the agents that IT deployed. But what about the Power Automate workflow that marketing built last month? The Azure AI project a developer spun up for a proof of concept? The ChatGPT browser window open on dozens of laptops across your organization? 

Shadow AI isn’t a theoretical problem. It’s happening right now, and most detection approaches miss more than they find. 

Why? Because Single-Method Detection Falls Short 

Most AI governance tools detect agents using one or two methods. They might scan Entra ID for service principals with AI-related permissions. Or they might monitor network traffic for calls to OpenAI’s API. And these approaches work… sort of. They find agents that happen to be visible through that particular lens. 

But here’s the thing about AI agents: they can exist in many different layers of your technology stack. An agent might authenticate through Entra ID, or it might not. It might make API calls that traverse your network, or it might run entirely within a SaaS platform. It might have a formal service principal, or it might be a script running on someone’s laptop. 

Any single detection method has blind spots. An agent that doesn’t use a service principal won’t show up in identity scans. An agent running inside Power Platform might not generate the network signatures your edge monitoring expects. A local Python script calling OpenAI might bypass every enterprise control except the one watching the endpoint. 

This is why finding AI agents requires thinking in layers, leveraging multiple complementary approaches that each catch what the others miss. 

The Six Planes of Detection 

Cyclotron Pulse approaches AI agent discovery through what we call the Six-Plane Detection Framework. Each plane represents a different layer of the technology stack where agents can be detected. Together, they create overlapping coverage that makes it nearly impossible for an agent to avoid all six checkpoints. 

How Planes Work Together

Individual detections are useful. Correlated detections across multiple planes are exponentially more powerful. Consider this scenario: your security team spots a Power App accessing customer data.  

Four planes independently identified the same agent. That’s not a maybe. That’s high-confidence detection with multiple evidence sources. The agent gets registered, its risk gets assessed, and governance policies apply automatically. 

Contrast that with single-plane detection. Maybe you’d catch the network traffic. Maybe you’d notice the DLP event. But without correlation, you’d have disconnected signals that might never get connected to a complete picture of what’s actually running. 

The correlation math works like this: three or more planes agreeing produces 80% confidence or higher. Four planes push confidence above 90%. Five or six planes detecting the same agent approaches certainty. 

Shadow AI and the Detection Gap

Shadow AI presents the hardest detection challenge. When employees use consumer AI tools through their browsers, there’s no service principal to monitor, no enterprise SDK to scan, and no formal deployment to intercept. 

Single-method detection typically misses these entirely. There’s nothing in Entra ID to find and there’s no code repository to scan, so from an identity perspective, nothing suspicious is happening. 

But shadow AI still leaves traces across multiple other planes. The Network plane catches the traffic to chat.openai.com, while the Data plane might notice the same user downloading confidential documents from SharePoint. And the Endpoint plane could detect clipboard activity copying between enterprise applications and browser windows. 

Any one of these signals might look innocent in isolation, but combined, they tell a story: someone is using consumer AI with company data. That’s exactly the kind of pattern multi-plane detection was built to catch. 

Organizations deploying six-plane detection consistently discover shadow AI usage they didn’t know existed. Sadly, finding 40-50 employees using unauthorized AI tools in the first week of deployment isn’t unusual. It’s typical. 

Coverage Reality Check 

Here’s a practical way to think about detection coverage. Single-plane detection finds around 60% of agents. It finds the obvious ones, the well-documented ones, the agents that happen to be visible through whatever method you’re using. Two or three planes push coverage to roughly 85%. You’re catching more edge cases, more unusual deployments, and more agents that don’t follow standard patterns. 

Six-plane detection reaches around 95%. Now, you’re even finding agents that exist only as code experiments. You’re catching infrastructure-based agents in Kubernetes clusters and identifying shadow AI that bypasses every official channel. The remaining 5% typically represents highly unusual configurations that can be addressed through targeted monitoring. 

The coverage difference translates directly to risk reduction. If you don’t know an agent exists, you can’t assess its risk, apply policies, or respond when it misbehaves. Every undiscovered agent is a blind spot in your security posture. 

Building Confidence Through Redundancy

The defense-in-depth philosophy behind six-plane detection assumes every method has weaknesses. An agent might evade one plane, maybe even two. Evading all six is exponentially harder. 

This redundancy creates confidence that auditors and compliance teams appreciate. When asked “how do you know you’ve found all AI agents?”, the answer isn’t “we checked the service principals” or “we monitored the network.” The answer is “we deployed six complementary detection methods, each catching what the others might miss, with correlation ensuring high-confidence identification.” 

That’s a different conversation than most organizations can have today. 

What Good Agent Detection Looks Like

Organizations with comprehensive AI agent visibility operate differently than those still guessing about their AI landscape. 

They can answer basic questions with confidence. How many AI agents do we have? What’s their risk distribution? Which ones access sensitive data? Who owns each agent? When did new agents appear? 

They can respond to incidents with complete information. When something goes wrong with an AI agent, they already know it exists, know its permissions, and know its behavior baseline. Investigation starts from context, not discovery. 

These organizations can demonstrate compliance with evidence. Auditors asking about their AI governance receive  documentation of comprehensive detection methods, not hand-wavy assurances about spot-checking. And they sleep better—or at least, the security team does. The nagging concern about unknown AI running somewhere in the environment fades when you’re confident about your detection coverage. 

Implementation Reality

Six-plane detection uses established monitoring approaches for each plane: SBOM scanning and Semgrep for code, Microsoft Graph API for identity, Cloudflare Workers for network edge, Kubernetes admission webhooks for platform, Purview integration for data, Defender queries for endpoints. 

The technical complexity isn’t in any individual plane, but in the correlation layer that connects detections across all six. That’s where agent identities get unified, confidence scores get calculated, and the complete picture emerges. 

Detection latency for new agents typically runs under five minutes. Correlation processing adds about fifteen seconds. False positive rates stay below 5%. These are production-grade numbers, not benchmarks from controlled environments. 

The Alternative: Hoping You Got Lucky

Without comprehensive detection, AI governance becomes a guessing game. You govern the agents you know about and hope there aren’t too many you don’t. That’s not a sustainable strategy as AI adoption accelerates. Employees are building agents. Developers are experimenting with AI frameworks. Third-party tools are embedding AI capabilities. The pace is only increasing. 

Every agent you don’t know about is a risk you can’t manage. Data access you can’t audit. Permissions you can’t review. Behavior you can’t monitor. Eventually, one of those unknown agents will cause a problem, and you’ll discover it existed only when something goes wrong. 

Ready to find out what’s actually running in your environment?

Cyclotron Pulse was built around six-plane detection specifically because comprehensive visibility is the foundation everything else depends on. You can’t score risk for agents you haven’t found. You can’t enforce policies on agents you don’t know exist. And you can’t govern and defend what you can’t see. The framework recognizes that AI agents can exist in many forms across many layers, and detection needs to match that reality. 

If you’re ready to find out what’s actually running in your environment, Cyclotron offers complimentary Pulse Risk Assessments for qualifying customers. We deploy six-plane detection against your infrastructure and show you exactly what we find, including the agents you didn’t even know you had. Because comprehensive AI agent governance starts with knowing what you’re governing.