Blog Posts
Most security teams don't know which AI tools employees are using. Polygraf AI presents a step-by-step guide to detecting shadow AI from network signals to endpoint data.
98% of organizations are using AI without permission – and most can't see it. Shadow AI lives inside approved browsers and sanctioned SaaS, where domain-blocking and app inventories miss it completely. This is the practical, how-to guide to actually find it: the detection layers, the method, and how to go from "we think we have a problem" to a full, governed view.
Every organization wants to know the answer to one question: "What AI tools are our people actually using?" It sounds like it should be a simple question to answer: pull a report, check the firewall logs, look at the SaaS inventory. It is not. Shadow AI is the hardest thing to detect because unlike the unauthorized Dropbox accounts of the shadow-IT era, AI usage often is hidden inside the tools you've already approved: a copilot in your sanctioned office suite, an AI feature in the CRM you pay for, a browser extension that quietly pipes data to an LLM. Domain-blocking and app inventories – the tools for shadow IT – just sail right over all of it.
Shadow AI detection needs a different approach: visibility of data flows and behavior, not just of application names. Here is the method. We will talk about why the obvious approaches are not enough, the detection layers that actually work and what each one detects, a step by step discovery process and how to go from one-time visibility to continuous governance.
Shadow IT was a set of discrete, unapproved apps – and CASB, SSO, and domain-blocking could find them. Shadow AI is hidden in approved browsers, sanctioned SaaS, and embedded features, so the prompt itself is the data path. You are not searching for an unapproved app, you are searching for sensitive data flowing into an AI feature in an approved app. That is why app-inventory tools leave most shadow AI hidden – and why detection must be at the data and behavior level.
Even when it is hiding, shadow AI is not invisible – it leaves traces in several layers. The problem is that each layer contains a different signal and none of them is complete by itself. The key to detection is to collect several and to correlate them.
| Signal | Where it appears | What it reveals |
|---|---|---|
| DNS / network traffic | Firewall, DNS resolver, proxy logs | Links to known AI service domains and API endpoints – the big first pass |
| Endpoint activity | Process, filesystem, local ports on the device | Local AI models, desktop AI apps, and agents that do not traverse any monitored network path. |
| Browser activity | Extensions, in-page interactions | AI browser extensions and AI features in web apps – the biggest blind spot. |
| OAuth grants | Microsoft 365, Google Workspace, SaaS | AI tools and extensions that you have been granted access to your core platforms, and that have been able to bypass access controls. |
| SaaS / API logs | Sanctioned platform admin consoles | "Approved platform, unapproved use" – AI agents and features within approved tools. |
There are five practical detection layers. The most important thing to know: each has a blind spot, and the blind spots are where the riskiest shadow AI lives. Hover each layer to see what it catches and where it fails.
Network and cloud-level detection (layers 1–3) tells you that *some* AI tool was called. The riskiest shadow AI – embedded copilots, browser extensions and the sensitive data going in them – is at the browser and content layers (4–5) where those methods are blind. This is why endpoint-level visibility is not one of many options, but the layer that fills the gap that the others structurally cannot.
With the layers known, here is the actual order to run. This works whether you are starting from zero or extending an existing program.
The biggest mistake is to use one method. No single layer will find all shadow AI, so run network, CASB/SaaS, OAuth and – most importantly – browser/endpoint detection at the same time and correlate the results. Each reveals tools the others do not.
Transform raw detections into a structured stock of all AI tools in use, who is using them, how often and how. Crucially, map activity back to specific departments and roles — operational accountability and targeted remediation both depend on knowing who, not just what.
Not all shadow AI is created equal. A marketer thinking of taglines is a different risk than a developer who pastes proprietary code or a clinician who enters patient data. Focus on the high-data-sensitivity, high-regulation intersections first.
Detection only works if it leads to a better path. Instead of silently blocking, redirect to approved tools with the same functionality and tell them why the approved option exists. People use shadow AI because it works, if the approved path is as fast, they will take it.
A one-off audit is stale the week after you do it – new tools are coming out all the time. Detection has to be continuous and visibility has to become control: the ability to look, redact or block sensitive data the instant it is about to go into an AI tool, no matter which one. That is the step that makes a report into protection.
Shadow AI lives in the approved browsers, SaaS, and embedded features. The prompt is the data path – that's why finding it requires seeing the data and content, not just a list of apps.
— Polygraf AI, on why shadow AI detection has to happen at the point of useFinding shadow AI is necessary but not enough. A list of tools doesn't protect anything – what protects you is the ability to act on what you find when it matters. This is the model of Polygraf AI's Desktop Overlay: three stages to turn raw visibility into durable control.
How much shadow AI risk are you carrying? Polygraf's AI Risk Calculator models your company's exposure (data breach, regulatory and litigation risk) and shows which obligations apply to you based on your industry, tools, data types and existing controls.
Use this to check if your detection program actually fills the gaps. If you can't check most of these, the riskiest shadow AI is probably still out of sight.
Polygraf's Desktop Overlay grabs the embedded copilots, browser extensions and local AI that network and cloud tools miss – and it controls what data gets to them, in-line. View, Control, Enforce. On-premise, sub-100ms, zero data egress.
At Polygraf, we envision a future where AI augments human capabilities without compromising safety, privacy, or ethical standards. Trust in our commitment to building this future with you.
© 2026 Polygraf AI. All rights reserved.
Our Solutions 
Regulations Your download will start now.
Please provide information below and we will send you a link to download the white paper.
Data Privacy 
Data Provenance 
Content Generation