AI Data Loss Prevention (AI DLP): How to Stop Data Leaks to ChatGPT and Other AI

Why classic DLP misses AI

Traditional DLP products were built to watch files and channels: email gateways, endpoints, cloud storage, network egress. They classify documents and block a file from being attached or uploaded. AI changes the shape of the problem in three ways.

• The data is typed or pasted, not attached. A prompt is free text in a web form. There is no file to fingerprint, and the sensitive part is often a few characters inside a long, harmless message.
• The destination looks legitimate. chatgpt.com, claude.ai and gemini.google.com are business tools your people are encouraged to use. Blocking them outright pushes usage to personal phones and unmanaged browsers, where you see nothing.
• Context matters. "Summarize this contract" is fine; "summarize this contract" followed by a client's bank details is not. Useful AI DLP has to read the message at the moment of sending and judge what is inside it.

What AI DLP should catch

Start with the categories that cause the most damage when they leak and are the easiest to detect reliably:

• Secrets and credentials: API keys, access tokens, private keys, database connection strings, passwords written into a message.
• Personal data (PII): names together with contact details, national ID and passport numbers, dates of birth, home addresses.
• Financial data: payment card numbers, bank account and routing numbers, salary information.
• Health data (PHI): diagnoses, medication, patient identifiers.
• Confidential business data: customer lists, unreleased financials, contracts, source code, internal hostnames.

Where to put the control

There are four places you can enforce AI DLP. Most organizations end up combining two of them.

• At the Send button (browser extension or app). The check runs where the person is typing, before anything leaves the device. It sees the exact text and attachments, can show what was found, and can offer to redact instead of simply blocking. This is how EraseAI's AI firewall works in Chrome and on Android.
• At the network (secure web gateway or CASB). Inspects traffic to AI sites. Good for visibility and for blocking unapproved AI apps, but it needs TLS inspection and struggles with the free-text, in-page nature of prompts.
• At the API layer. For AI features your own developers build, scan inputs and outputs in code before they reach a model. EraseAI's API does this for teams building on LLMs.
• At the provider. Enterprise AI plans add retention controls, no-training commitments and audit logs. These reduce what happens to data after it is sent; they do not stop it being sent.

A practical AI DLP rollout

1. Find out what is in use. List the AI tools people already use, approved or not. Ask; don't only rely on logs. This is your shadow AI inventory.
2. Write a short policy. Which tools are approved, which data must never go into any AI tool, and what to do when in doubt. Our generative AI acceptable use policy template is a starting point.
3. Put a check at the Send button on managed browsers and phones, so mistakes are caught when they happen rather than found in an audit months later.
4. Prefer redaction to blocking. If the tool replaces a card number with [CARD] and lets the question through, people keep using the approved tool instead of working around you.
5. Review what is being caught. Patterns in the findings tell you where training, templates or process changes are needed.

AI DLP for individuals

You don't need a security team to protect yourself. Turn off model training on your chats where the provider allows it, don't paste whole documents when a summary of the relevant part will do, and use a tool that checks each message before it is sent. EraseAI's Chrome extension is free and checks every message on your device, with no account needed.

Frequently asked questions

Related guides