How to stop staff pasting sensitive data into ChatGPT and other AI tools
Someone in your organisation pasted something into an AI tool today that should not have left. Usually it is a customer list to tidy up, a contract to summarise or code to debug, done to get the job finished rather than to cause harm. Bans, policies and data loss prevention rules each catch part of it. This guide sets out where each one leaves a gap, and why we recommend Mimecast Incydr to close it. C4C is an accredited Mimecast partner that supplies and supports Incydr, disclosed up front.
Most AI data leakage is a paste, not a file. Mimecast reports that around 90 percent of shadow AI data loss happens through paste events and 10 percent through file uploads. Source: Mimecast, shadow AI. You cannot stop what you cannot see, and you cannot write a rule for every kind of sensitive text. The control that works sees every paste and upload into AI tools, knows who did it, and blocks the unsanctioned destinations while leaving the approved tool open.
Pasting into AI is the new version of emailing a file to a personal account, except it is quicker, feels harmless and leaves no trace in the systems most security teams watch. Our guide to securing enterprise AI covers the wider risk, including Microsoft Copilot and oversharing. This guide answers the narrower, practical question: how do you actually stop sensitive data being pasted into ChatGPT, Claude, Gemini and the rest?
Why the usual answers leave a gap
Most organisations reach for one of five responses. Each has a place. None closes the gap on its own.
- Block AI sites at the firewall. People switch to their phones, home machines and personal accounts, and you lose what little visibility you had. New AI tools appear every month, so a block list is always behind, and a blanket block also stops the use you want to encourage.
- Write an acceptable use policy and train people. Necessary, and it sets the rules. But a policy is not a control. It cannot tell you whether anyone is following it, and the people who paste a client list in a hurry usually know the policy perfectly well.
- Buy an enterprise AI tool. The single best first step, because a sanctioned tool with contractual data protection removes most of the reason to reach for a public one. But it does not stop someone opening a personal ChatGPT account in the next browser tab, and the contract only protects the tool you bought.
- Use Microsoft Purview data loss prevention. Purview endpoint DLP can inspect text as it is pasted into a supported browser and audit, warn or block when it matches the sensitive information types you have defined. That is a genuine capability, and it works well for structured data such as card numbers or National Insurance numbers. Its limit is that it only catches what your rules describe. A pricing strategy, a board paper, a contract clause or a block of source code rarely matches a sensitive information type unless someone has built and tuned a classifier for it. It also needs devices onboarded to Purview, the right Microsoft 365 licensing, and an extension for Chrome and Firefox on Windows. Source: Microsoft Learn, restricting paste into browsers.
- Filter web traffic with a secure web gateway. Network tools can allow or block AI sites by category, and some inspect content. They see traffic, not people. They cannot tell a leaver pasting a client list from a developer pasting a public code sample, content inspection depends on decrypting the traffic, and devices off the corporate network depend on how the agent is deployed.
The common thread is that each answer either cannot see the paste, cannot judge it without a rule, or cannot tell who did it and why that matters.
What each approach can actually see
| What matters | Block AI sites | Microsoft Purview DLP | Mimecast Incydr |
|---|---|---|---|
| Sees pastes into AI tools | No. It blocks the site, or sees nothing. | Yes, on onboarded devices in supported browsers. | Yes, in Chrome, Edge and Firefox on Windows and Mac. |
| Judges whether the content is sensitive | No. | Yes, where it matches a sensitive information type you have defined. | Verifies the content of files uploaded to AI tools. For pastes it records the event, the tool and the person, and controls the destination. |
| Works without writing rules first | Yes, but only as a blanket block. | No. The policies are the product. | Yes. It records activity from day one and scores the risk. |
| Keeps the sanctioned tool open | Only with careful allow lists. | Yes, through domain groups in the policy. | Yes. Block pastes and uploads to untrusted destinations, allow the approved one. |
| Knows who did it and their context | No. | Alert detail on the user and the rule matched. | Yes. Tied to the person, their role and their recent activity, with leavers scored higher. |
| Sees the other routes out | No. | Within the Microsoft estate and onboarded devices. | Yes. Cloud sync, personal cloud, removable media, Git push and pull, AirDrop and more. |
| Time to useful | Immediate, and immediately worked around. | Weeks to months of classification and tuning. | Days. Baseline first, then act on the exceptions. |
Incydr capabilities from Mimecast insider risk management and Mimecast, March 2026. Purview capabilities from Microsoft Learn.
Why we recommend Mimecast Incydr for this
The paste problem is a visibility problem before it is a policy problem, and that is why Incydr is the product we recommend. Four reasons, specifically.
- Control the destination, not the content. You cannot write a rule that recognises every sensitive paragraph, but you do know which AI tools are approved. Incydr can block pastes and uploads to untrusted destinations while your sanctioned tool stays open, so the protection does not depend on classifying the text correctly.
- It knows who, not just what. The same paste means something different from a developer on a normal day and from someone who resigned last week. Incydr ties every event to the person and their wider pattern of activity, scores the risk, and puts the cases that matter in front of security or HR rather than burying them in alerts.
- It sees the whole route out, not just AI. AI is one exit among many. The same agent sees cloud sync, personal cloud accounts, removable media, Git activity and AirDrop, across Windows and Mac, and Microsoft 365, Google Drive and Box, so you are not buying a point tool for each new channel.
- It is useful in days. There is no classification project before you see anything. Run it in monitor mode, see what is really being pasted and where, then switch on blocking for the destinations that warrant it.
The approach is proven. For a leading UK financial services organisation with strong perimeter security and no insight into how its people moved data, C4C deployed Incydr with custom risk scoring and automated triage. Within ninety days it had full visibility of data movement across endpoints and cloud and an 80 percent reduction in unmonitored file transfers. The Incydr case study has the detail.
If your estate is Microsoft end to end, you already own the licensing, your sensitive data fits clear information types such as card or account numbers, and a compliance team will write and tune the policies, Purview paste controls may be enough for this problem. We will say so. Our Incydr vs Microsoft Purview comparison goes through the trade off in full.
The order that works
- Give people a sanctioned tool. An enterprise AI plan where your data is contractually excluded from training removes most of the reason to paste into a public one. Our Copilot vs ChatGPT vs Gemini vs Claude comparison helps choose.
- See what is really happening. Deploy Incydr in monitor mode for a few weeks. Most organisations are surprised by which tools are in use, which teams use them, and what goes in.
- Block the unsanctioned destinations. Turn on blocking for pastes and uploads to untrusted AI tools, keep the approved one open, and handle the high risk cases, such as leavers, with a proportionate response.
- Update the policy from the evidence. Write the acceptable use policy around what you found, and train on real examples rather than hypotheticals.
Who we are. C4C Group is an independent UK technology consultancy and an accredited Mimecast partner. We supply, implement and support Mimecast Incydr, and this guide recommends it deliberately. We also help organisations choose and buy the enterprise AI tool that removes the reason to paste in the first place, and we will tell you if what you already own is enough. Ask us what Incydr would show in your organisation.
Find out what your staff are pasting into AI
Most organisations have no idea which AI tools are in use or what is going into them. We will show you, then help you stop the risky part without stopping the useful part.
Prefer email? Reach us directly at hello@c4cgroup.co.uk.
Frequently asked questions
Can you stop employees pasting sensitive data into ChatGPT?
Yes, but not with a ban or a policy on its own. Blocking AI sites pushes people to phones and personal accounts, and a policy cannot show you whether it is being followed. What works is a sanctioned enterprise AI tool so people have a safe option, plus a control on the endpoint and browser that sees pastes and uploads into AI tools and can block them for unsanctioned ones. Mimecast Incydr does this in Chrome, Edge and Firefox on Windows and Mac.
Does ChatGPT train on what staff paste into it?
It depends on the plan. With consumer tools, what people enter can be retained and, depending on the tier and the user settings, used to improve the models. OpenAI commits not to train on business data from ChatGPT Enterprise and its business plans by default. The risk is that staff use a personal account in another browser tab, where your organisation has no contract, no settings and no record.
Is blocking ChatGPT at the firewall enough?
No. People move to their phones, home machines and personal accounts, and you lose the little visibility you had. There are also dozens of AI tools, with new ones every month, so a block list is always behind. Blocking also stops the legitimate use you want to encourage. Control the destination per tool, keep the sanctioned one open, and see what happens on the endpoint.
Can Microsoft Purview stop staff pasting into AI tools?
Partly. Purview endpoint DLP can inspect text as it is pasted into a supported browser and audit, warn or block when it matches the sensitive information types you have defined, such as card or National Insurance numbers. It needs devices onboarded to Purview, the right Microsoft 365 licensing and well tuned policies. It is weaker on data no rule describes, such as a pricing strategy, a contract clause or source code, and on telling a leaver from a careless user.
How does Mimecast Incydr detect data pasted into AI tools?
An agent on Windows and Mac and an extension in Chrome, Edge and Firefox record paste events and file uploads into browser based AI tools such as ChatGPT, Claude, Gemini and DeepSeek, tied to the person who did it. Incydr can verify the content of files uploaded to an AI tool, block pastes and uploads to untrusted destinations, and scores each event against the user and the wider pattern of their activity, so the risky cases rise to the top.
Is C4C independent if it recommends Mimecast Incydr?
C4C is an accredited Mimecast partner and supplies, implements and supports Incydr, and we say so on every page that recommends it. This guide is deliberately vendor forward. We still tell clients when the tools they already own are enough, for example a well governed Microsoft estate whose sensitive data fits clear information types.