Shadow AI Data Leakage: How to Find the AI Tools You Didn't Approve and What They've Been Sent

Shadow AI data leakage is company data sent to AI tools that IT never reviewed or approved. It's harder to stop than leakage through sanctioned tools because you don't know where to look. The fix is a sequence: discover every AI app in use, attribute it to people and account types, inspect what data moved, then enforce a policy. DLP on its own isn't enough, because you can't write a policy for a tool you don't know exists.

Your approved AI list has three tools on it. Your network traffic has thirty.

Key takeaways

• Shadow AI is any AI tool in use without IT's approval, including approved tools used with personal accounts.
• Discovery comes first. You can't protect data going to tools you can't see.
• Attribution matters. Knowing an app is in use is less useful than knowing who uses it, how much, and with which account.
• DLP shows what moved. Volume alone doesn't tell you whether it was sensitive.
• Enforcement should take one step, from the same screen where you found the app.

What counts as shadow AI?

More than most lists assume.

• Unapproved chatbots and models. The new one everyone's trying this month.
• AI coding assistants installed by individual engineers.
• Meeting note-takers that join calls and transcribe them.
• AI features inside other SaaS apps that were switched on without review.
• Approved tools on personal accounts. ChatGPT is sanctioned. A personal ChatGPT login isn't.

That last one is the most common and the easiest to overlook, because the domain is on your allow list.

Why is shadow AI a data problem and not just an IT problem?

Unapproved software used to mean a license or support headache. With AI, the tool's whole function is to take in content.

• You have no agreement with the vendor. No data handling terms, no retention commitment.
• You have no admin view. No audit log, no way to see or delete what was submitted.
• You have no account control. When the employee leaves, the account and its history go with them.

Why can't most tools see it?

• DNS filtering sees domains, not accounts or content. It can't tell an enterprise tenant from a personal login.
• Surveys rely on memory and honesty. People under-report, and the list is stale in a month.
• SaaS admin consoles only show the tools you already manage.
• Legacy DLP needs a destination and a rule. It can't inspect traffic to a tool nobody told it about.

You need something that sees AI traffic from the device, per user, with the account attached.

The four-step method: discover, attribute, inspect, enforce

1. Discover

AI Analytics in dope.security lists every AI app seen on your endpoints, sanctioned or not. No agent survey, no proxy logs to grep. You see ChatGPT, Claude, Gemini, and Copilot next to the long tail: coding assistants, note-takers, and models you've never heard of.

2. Attribute

For each app, you get per-user transactions and data volume, and whether the account is personal or enterprise-licensed. Click an app to see its top users.

3. Inspect

Dopamine DLP reads prompts and attachments, so you see what data actually moved, not just how much. Text is extracted on the device and classified by an LLM with zero data retention. Detections are sorted into PII, PCI, PHI, and intellectual property.

4. Enforce

Block the app from the AI Usage screen, and the policy is live on every endpoint in seconds. Or allow the enterprise tenant and block personal accounts with Cloud Application Control.

Ask instead of digging

The Dopamine agent in dope.console answers plain-language questions over your analytics. Three that matter for shadow AI:

• "Top AI apps and domains." Who's using what over the last seven days.
• "Sensitive data sent to AI." Source code, PII, and secrets, by user.
• "Users to investigate first." Ranked by blocks and DLP violations, with the reasoning attached.

Answers come back in under 10 seconds and export to CSV in one click. One example from the product: "Is anyone in the company using DeepSeek?" returned the user and the count in 7 seconds.

What should you do with each app you find?

That last row matters. Shadow AI is also a demand signal.

A 30-day shadow AI cleanup

1. Week one. Deploy the agent through MDM. Let AI Analytics collect a full week.
2. Week two. Review the list. Classify each app as sanctioned, tolerated, or not allowed. Run DLP in Monitor.
3. Week three. Enable Cloud Application Control for sanctioned tools. Block the not-allowed list. Set tolerated tools to Warn.
4. Week four. Move DLP to Block. Share a short summary with leadership: apps found, personal-account share, sensitive detections, actions taken.

Then make the weekly review a habit. Fifteen minutes with the three questions above is enough.

What to tell employees

Blocking without explanation breeds workarounds. Tell people three things:

• Which AI tools are approved, and how to get access.
• That personal accounts aren't allowed for work, and why.
• What kinds of data should never go into any AI tool.

Then put the same message on the Warn and Block screens, so it shows up at the moment it's needed.

Limits to know

dope.security sees AI use on managed Mac and Windows devices. It doesn't see personal phones or unmanaged laptops. DLP coverage for prompts includes ChatGPT, Claude, Perplexity, Abacus, and Copilot, plus file uploads to the web. For other AI apps, discovery and access control still apply.

Frequently asked questions

What is shadow AI?

AI tools used for work without IT's review or approval, including approved tools used through personal accounts.

What is shadow AI data leakage?

Company data being sent to those unapproved AI tools or accounts, outside your agreements and visibility.

How do I discover shadow AI in my company?

Use a tool that observes AI traffic from each device and attributes it to users and accounts. AI Analytics in dope.security does this without surveys.

Can DLP stop shadow AI?

DLP stops sensitive content from being sent. You also need discovery to find the tools and access control to restrict them.

Should I block every unapproved AI tool?

Not automatically. Block the risky ones, warn on the rest, and review popular tools for approval.

How often should I review AI usage?

Weekly works well. New tools appear constantly, and a short review catches them early.

Find out what's running

Start a free trial, see how teams manage AI, or read about the best DLP for AI.