How to Stop Employees Leaking Sensitive Data into ChatGPT
Your team is already using AI tools at work. Some of that use is sanctioned, plenty of it isn't, and almost none of it is visible to the people responsible for protecting company data.
That isn't because employees are careless. It's because ChatGPT, Claude and Gemini are genuinely useful, and the fastest way to get a useful answer is to paste in the real thing: the real customer ticket, the real config file, the real board deck. The risk is in that paste.
This guide covers why the obvious fix (banning AI) tends to fail, what actually leaks, and a five-step approach that lets people keep using AI while sensitive data stays inside the business.
Why banning AI tools rarely works
In April 2023, Samsung engineers pasted internal source code into ChatGPT while trying to fix a problem. The following month, Samsung restricted employee use of generative AI tools on company devices.
Bans like this are an understandable first reaction, but they come with two problems:
- People route around them. Blocking one domain doesn't stop personal accounts on personal phones, browser extensions with built-in AI, or the next new tool that isn't on the block list yet.
- You lose the productivity. The competitors who work out how to use AI safely move faster than the ones who simply switch it off.
A ban also leaves you just as blind as before. You stop seeing the usage on your own network, but the usage doesn't stop.
What actually gets pasted into AI tools
The leaks are rarely dramatic. They're everyday tasks done with real data:
- Customer support: an agent pastes a full ticket, with the customer's name, email, phone number and account number, and asks for a polite reply.
- Engineering: a developer pastes a config file to debug a failing login, with the production database URL and API key still in it.
- Finance: an analyst uploads board materials and asks for an executive summary, including revenue forecasts and deal pipeline.
- HR: a manager drafts a performance review or termination letter using an employee's name, salary and medical leave details.
- Legal and professional services: client names, case details and privileged communications go in for drafting or analysis.
In each case the employee is trying to do their job well. The problem is that the data has now left the company, gone to a third party, and may be kept according to that provider's retention settings rather than yours.
Why your existing DLP doesn't catch it
Most data loss prevention tools were built to watch email, file shares and USB drives. AI usage looks different:
- The data is typed or pasted into a web page, not attached to an email or copied to a device.
- It travels over an encrypted connection to a legitimate, often approved, domain, so network tools see normal web traffic.
- It happens in seconds, inside the browser, often from a personal account the company doesn't manage.
That's why most organizations find out about AI data exposure from an audit or an incident, not from a tool.
A five-step approach that works
1. Find out what's actually in use
Before writing policy, get an honest picture: which AI tools people use, which teams use them most, and what kinds of tasks they're doing. You'll almost always find more tools than you expected, and a handful of use cases that are clearly worth supporting.
2. Classify your data, then decide an action for each type
"AI: yes or no" is the wrong question. The better question is "which data, in which tool, with what action?" A practical model uses four actions:
| Action | When to use it | Example |
|---|---|---|
| Allow | Low-risk or public information | Public company info, general questions |
| Warn | Context matters; let the user decide with a reminder | Internal strategy, product roadmaps |
| Mask | The task works fine without the real values | Customer names and emails in a support reply |
| Block | Never acceptable to send | Card numbers, credentials, bank details |
Masking is the one most teams overlook. A support agent asking for a reply template doesn't need the real customer's details in the prompt. Swap them for placeholders and they still get the help they wanted.
3. Write a short, plain-language AI policy
Keep it to a page. Say which tools are approved, which kinds of data must never go into any AI tool, and who to ask when unsure. A policy people actually read beats a thorough one nobody opens.
4. Put the control where the data leaves: the browser
Training helps, but people are busy and a paste takes a second. The most reliable control sits at the point where the prompt is about to be sent, inspects it in real time, and applies your policy before anything leaves the browser. Done well, this adds no noticeable delay and doesn't require routing traffic through a proxy.
A warning that appears at the moment of risk, with the reason ("this looks like a customer email address; company policy is to mask these"), also teaches people far better than an annual training module.
5. Keep an audit trail
When an auditor, a customer's security questionnaire or a regulator asks how you control AI use, "we have a policy" isn't a strong answer. A log of what was detected, in which tool, by whom, and what action was taken, is. It also shows you where policy needs tuning: if one team triggers constant warnings, they may need an approved tool rather than more reminders.
A quick checklist
- We know which AI tools our employees use, including on personal accounts.
- We've decided which data types are allowed, warned, masked or blocked.
- We have a one-page AI use policy that people have actually seen.
- Sensitive data is checked before it leaves the browser, not after.
- We can show an auditor a record of AI-related data incidents and how they were handled.
If you can't tick most of these yet, you're in good company. Most organizations are still at step one.
Where Nura fits
Nura is a browser extension that does steps 4 and 5 for you. It inspects every prompt sent to ChatGPT, Claude, Gemini, Copilot, Perplexity and DeepSeek, detects sensitive data across 47 categories out of the box, and applies your Allow, Warn, Mask or Block policy per data type, per team and per AI tool. Every decision is logged for compliance and audit.
Your team keeps the AI tools that make them faster. Your data stays where it belongs.