How can you prevent ChatGPT from becoming your worst enemy in cybersecurity?

Nearly 40% of companies using AI have already been targeted by cyberattacks specifically aimed at these tools. That statistic is striking. But what concerns me even more is what it doesn’t reveal: how many incidents go completely unnoticed, simply because no one realized there was a risk in the first place?

At CreativMinds, we support organizations of all sizes with their cybersecurity challenges. And since ChatGPT has become part of everyday workflows, one question keeps coming up: “We can use it, but… is it really safe?”

The short answer: it depends. The long answer is what this article is all about.

The real problem isn’t the tool itself. It’s how we use it

ChatGPT isn’t inherently dangerous. What is dangerous is the gap between what people think it does and what it actually does.

Let’s take a concrete example. Someone in HR uses ChatGPT to rephrase a candidate rejection email. So far, nothing problematic. But to “save time,” they copy and paste the candidate’s CV into the prompt — name, address, phone number, full professional history. That data has now been processed through OpenAI’s servers.

Does that mean it will be reused or exposed? Not necessarily. But it has left the company’s controlled environment — and that’s something most employees don’t fully realize.

Another common scenario: a salesperson asks ChatGPT to “draft a commercial proposal for client X,” including negotiated pricing, specific terms, and sometimes even margin details. Strategic information, shared in what seems like a harmless request.

The three risks we (almost) always underestimate

1. Unintentional data leaks

This is the most common and the least obvious risk. No one sets out thinking, “I’m going to leak confidential data.” It happens out of convenience, habit, or simply a lack of awareness.

The issue is that ChatGPT is designed to be helpful. The more context you provide, the better the response. Naturally, this encourages users to share more than they should.

What you can do: Before typing anything, ask yourself a simple question: “Would I send this information to a complete stranger by email?” If the answer is no, it doesn’t belong in ChatGPT either.

2. AI-assisted social engineering

This one is more insidious. Malicious actors are using tools like ChatGPT to craft highly convincing phishing emails. Gone are the obvious spelling mistakes and clumsy phrasing that used to raise red flags.

I recently came across an AI-generated email that perfectly mimicked the tone of an IT director requesting an “urgent access review.” Even a cautious employee could have fallen for it.

On the flip side, a compromised or poorly configured internal chatbot can itself become an attack vector. Imagine a “technical support” assistant asking for your credentials using exactly the right tone, the right wording — and responding convincingly to your follow-up questions.

What you can do: Train your teams to recognize these new types of attacks. Traditional anti-phishing reflexes are no longer enough. Two-factor authentication should be non-negotiable for any sensitive action.

3. Regulatory compliance

GDPR, Switzerland’s FADP, industry-specific regulations… As soon as ChatGPT is used with personal or sensitive data, you’re stepping into a complex regulatory landscape.

The challenge is that many organizations simply haven’t mapped these use cases. ChatGPT has quietly made its way into day-to-day workflows without going through any kind of legal validation. And when an audit or an incident occurs, the full extent of the issue suddenly comes to light.

What you can do: Start with a clear assessment. Who is using what, for what purpose, and with which data? Without that visibility, securing anything is impossible.

What actually works (beyond good intentions)

Set clear — but realistic — guidelines

Banning ChatGPT doesn’t work. Employees will find workarounds, and you’ll lose all visibility into how it’s actually being used.

What works better is clearly defining what’s allowed, what isn’t — and most importantly, why. A framework that people understand is one they follow. One that’s imposed without explanation will be bypassed as soon as possible.

In practice, it can look like this:

Allowed: rephrasing generic content, brainstorming, researching public information
Prohibited: any personal data, any client information, any non-public financial data
Grey area: when in doubt, ask first

Implement technical safeguards

For organizations looking to go further, there are solutions available. Some can analyze prompts before they’re sent to the AI and automatically block those containing sensitive data. Others offer “enterprise” versions with stronger guarantees around data handling.

But be careful: technology is not a substitute for awareness. A poorly configured filter creates a false sense of security — and no tool can anticipate every possible scenario.

Train, and keep training

I know — “training” doesn’t exactly get people excited. But it’s what makes the difference between an organization that suffers from risks and one that manages them.

What works best, in our experience: short sessions built around real-life scenarios, with hands-on exercises. Not 45-slide PowerPoint presentations that are forgotten the moment people leave the room.

And above all: repeat the message. Cybersecurity isn’t a once-a-year training session — it’s a culture that’s built over time.

The specific case of companies building with AI

For organizations integrating ChatGPT (or other models) into their own products or processes, the stakes are even higher.

A few key points to watch:

• The API is not the public version. Data handling terms are different — and they need to be read carefully (really carefully).
• Training data. If you fine-tune a model using your own data, where does that data go? Who has access to it? For how long?
• Security testing. Regular audits of AI integrations should be part of the development lifecycle, just like functional testing.

What I’ve learned from supporting companies on this topic

Most of the ChatGPT-related incidents we’ve observed weren’t sophisticated attacks. They were human errors — made by capable people who simply weren’t aware of the risks.

What also stands out is the gap between how quickly these tools are adopted and how slowly organizations adapt their security practices. ChatGPT made its way into companies within weeks. Security policies, on the other hand, often evolve on annual cycles.

There’s a gap to close — and the longer you wait, the wider it gets.

Where should you start?

If you haven’t put anything in place yet, here’s a realistic way to get started:

Week 1: Run an informal assessment. Simply ask your teams how they’re using ChatGPT. No judgment — just aim to understand.

Weeks 2–3: Draft simple guidelines. One page max. What’s allowed, what isn’t, and who to contact if there’s any doubt.

Month 2: Organize an awareness session. Not a formal training — more of an open discussion with real-world examples.

Month 3 and beyond: Assess whether technical solutions are needed based on your context. And plan regular reminders.

The questions we get asked most often

Does ChatGPT store the data you send it?

By default, yes — at least temporarily. OpenAI states that conversations may be used to improve its models, unless you opt out in the settings or use the API/Enterprise versions. In practice, assume that anything you enter can be read and stored. It’s not a confidential space.

Can using ChatGPT put us at risk of GDPR non-compliance?

Potentially, yes. If you input personal data (names, emails, client information, etc.) without a valid legal basis or without informing the individuals concerned, you’re entering risky territory. Data transfers to the United States add another layer of complexity. For professional use involving personal data, it’s safer to rely on solutions with appropriate contractual safeguards.

What’s the difference between ChatGPT Free, Plus, Teams, and Enterprise?

Beyond features, the main difference lies in data handling. Free and Plus versions may use your conversations for training (unless you opt out). Teams and Enterprise offer stronger guarantees: no use of your data for training, enhanced encryption, and — for Enterprise — more controlled deployment options. If your company relies on ChatGPT regularly, upgrading to a business plan is worth considering.

How can I tell if my employees are using ChatGPT?

Honestly? You can’t fully control it — especially with personal devices and remote work. The best approach isn’t surveillance, but transparency: ask openly, without judgment. You’ll likely be surprised by how widely it’s used. And that visibility is exactly what allows you to put proper guidelines in place.

Can ChatGPT be used to process health or financial data?

This is highly sensitive. Such data is subject to strict regulations (medical confidentiality, banking regulations, etc.) requiring high levels of protection. In most cases, using the public version of ChatGPT with this type of data is not compliant. Specialized solutions exist for these sectors, with certified hosting and appropriate safeguards.

Are ChatGPT’s built-in privacy safeguards sufficient?

OpenAI has implemented certain safeguards, but they’re not foolproof. The model may refuse to generate some sensitive content, but it won’t automatically detect that you’re sharing confidential information. The responsibility remains with the user. Don’t rely on the tool to protect you from yourself.

What should you do if an employee has already shared sensitive data?

Don’t panic — but don’t downplay it either. Document what happened: what data was shared, when, and in what context. Depending on the nature of the information (personal data, trade secrets), you may have notification obligations. It’s also an opportunity to reinforce best practices across the team, without singling anyone out. These mistakes are often systemic before they are individual.

Are alternatives to ChatGPT (Claude, Mistral, etc.) safer?

Each tool has its own data policies. Mistral, for example, offers European hosting options that can simplify GDPR compliance. Claude (Anthropic) has different commitments regarding data usage. But “safer” largely depends on your use case and context. The key is to read the terms of service — really read them — and make an informed decision.

ChatGPT can be a powerful ally for productivity. But like any powerful tool, it requires a certain level of control to avoid backfiring.

The good news is that the risks are manageable — as long as you don’t pretend they don’t exist.