Yazen Rahmen, a Cybersecurity expert at SearchInform, explains how pasting confidential data into ChatGPT triggers instant compliance exposure and irreversible loss of control, and why visibility—not bans—is essential to managing AI‑driven leak risks.
Summarizing a hundred-page report, structuring a messy dataset, drafting boilerplate text – GenAI does all of this faster than a person can. Developers who have spent a week untangling AI-generated code may disagree, but the trend itself is settled. ChatGPT, Gemini, DeepSeek and Copilot are now part of everyday workflows, in large enterprises and small businesses alike. For all their undeniable benefits, these tools bring new risks – and the biggest one is that AI services are turning into yet another channel for corporate data leaks.
In Q2 2025, Harmonic Security analyzed one million prompts and 20,000 files sent to roughly 300 GenAI services. Sensitive information was found in 4.37% of prompts and in 22% of uploaded files. Most often, employees shared legal documents and source code.
Now imagine confidential company data has been uploaded to ChatGPT by an employee. What could follow, and what happens to that information?
Regulatory Claims
The first concern for management is usually the regulator. When an employee pastes a customer list into a private ChatGPT session to finish a report faster, several claims become possible at once: processing beyond the stated purpose, disclosure to a third party without a lawful basis, cross-border transfer without the required safeguards, and failure to apply adequate technical and organizational measures.
AI Doesn’t Keep Secrets
There’s another, often overlooked problem: once information is uploaded to an AI service, the company loses control over it entirely.
Samsung’s semiconductor division produced one of the first widely reported examples, in April 2023. An engineer working on an internal database ran into a problem with source code and asked ChatGPT to fix it. Within weeks, two further incidents followed in the same division: another engineer submitted proprietary code for optimization, and a third employee uploaded a recording of an internal meeting to have the minutes written up. Confidential material was now held on a third party’s servers, outside the company’s own retention and deletion controls, with no way to pull it back. Samsung restricted the use of generative AI tools on corporate devices shortly afterwards.
Technology that took years of work and considerable investment to develop can be compromised by a single attempt to save an afternoon. It is an expensive way to lose a competitive edge.
On top of that, AI services may use the content of user conversations to train their models. OpenAI, for example, states directly that conversation data is used for training unless the user opts out in the settings.
Does this mean corporate secrets uploaded to a chatbot are guaranteed to surface in answers to other users? Probably not. But it would be wrong to claim the risk doesn’t exist at all. A model may memorize individual fragments of uploaded code, personal data, or commercial information and reproduce them in its responses. It’s also worth remembering that, in certain situations, the AI provider’s employees and contractors responsible for security and quality control may have access to this data.
Does Banning AI Work?
So how can leaks like these be prevented? Ban AI altogether?
People take the path of least resistance, and prohibiting something that makes work easier simply moves that work onto personal devices and personal accounts, where the security team can see none of it. The exposure does not disappear; it stops being visible.
The most sensible way to minimize AI-related leak risks is not a total ban, but visibility and control.
AI platforms may look like brand-new tools, but in essence they are familiar, long-known leak channels. Employees upload data to chatbots either through a browser or a client application. That means a solution capable of monitoring web traffic, applications, files, and copy-paste operations can significantly reduce the risk of leaks through AI.
What a DLP System Needs to Handle
Modern DLP systems, which many companies have already adopted, provide solid control over all of these channels – which makes them highly relevant in the AI era as well.
Let’s look at the capabilities a DLP system needs in order to protect a company from leaks via ChatGPT and other AI services as effectively as possible.
- Deep content analysis
Catching a paragraph copied word for word into a chat window is trivial. The harder case, and the far more common one, is the fragment an employee has reworded before pasting it, or the document uploaded with a request to “rewrite this as a summary.” A DLP system has to recognize the sensitive original inside the modified version and stop the transfer before it leaves the perimeter.
- Policies out of the box, and policies that can be built
The data requiring protection is broadly the same everywhere: identity documents, banking details, medical records, contracts. A mature product ships with an extensive library of preconfigured policies covering these. It also has to let security teams write policies around whatever is specific to their business, whether that is drilling data, tender pricing or patient records.
- Blocking specific prompts and operations
Blocking a particular prompt, file type or operation allows a company to close the risky use cases while leaving the tool available for everything else. Employees learn quickly which tasks cannot be delegated to AI, and the security team avoids becoming the department that switched off the productivity tool.
The Bottom Line
Comprehensive control over AI use in a company – unlike a total ban – strikes a balance between security and efficiency. The business keeps the benefits of artificial intelligence while minimizing the risk of losing valuable information.