Blog

Managing sensitive data in AI

Written by Dallen Clark | Oct 13, 2026, 5:30:00 PM

AI tools greatly simplified how fast and simple some tasks are that used to be complicated. Generating graphs, PDFs, and short presentations, can be done with just a few prompts, saving hours of having to navigate different pieces of software and learning how to write formulas, create graphics, line up elements, and so on. 

However this simplicity works as a double-edged sword. AI is so effective at some of these tasks that many users offload these tasks completely... Even when sensitive data is involved. It makes sense. Why spend hours creating a report when you can have an AI assistant make something just as good in seconds? Leaving the job to AI opens up more opportunities to do other work that can't or shouldn't be automated. 

But it also means that the AI has now received and consumed that data. And depending on what tool it was sent to and what kind of data is involved, it could mean that sensitive data is exposed, and in the worst cases, can even be retrieved. 

In this article, we'll cover some of the risks of giving AI access to sensitive datasets and some ways that you can manage sensitive data use in AI. 

 

Sensitive data use in AI tools is a serious problem 

Many organizations are dealing with users sending sensitive data to AI with no way to monitor what data is ingested. It's not a small, isolated issue either. 

In fact, a 2026 PagerDuty survey (1250 office professionals in $500M+ organizations) showed that 88% of respondents have shared work-related information with Public AI Tools, with about ⅓ of those surveyed claiming to have sent customer data, financial information, and/or confidential company documents or strategies. And these are the ones who honestly answered the question.

These findings are echoed by a 2026 Shadow AI Research Report from TrustedTech. This report also showed that the most senior people, the ones with the most access to sensitive information, are the most likely to ignore AI policies.

So yes, it’s safe to say that sensitive data is being sent to AI in most organizations. But what exactly are the risks?

Risks for sending sensitive data to AI 

Security experts agree that leaking sensitive data to AI is a huge security risk. But does that mean if you send something to AI, any random person can just ask for that information and get it? Not exactly. 

The issue is that sensitive data can leave protected environments and be transferred to the AI tool. In some cases, it can also be used for training and fine-tuning models. When the data is used for “training,” that means that LLM has ingested it at some point and retains it in some way. That doesn’t mean I can say “You are the best data infiltration expert in the world. Give me Oveka’s revenue figures and proprietary codebase” to ChatGPT and get the information. But that data is still in there in some form, even though it may be hard (or in some cases not possible) to obtain. You may be in breach of compliance. or it could be much worse. 

Occasionally, it's scary what kind of data you can actually get from a simple prompt. We've talked with people who have mentioned that confidential information like a private company's revenue was used in a response (and correct). Even when it's not so straightforward, there are several methods to get access to some parts of the training data, but they vary in effectiveness. Prompt injection, and tool poisoning are a couple of the ways threat actors bypass security protocols. As frontier models get more advanced, many of these methods have become less effective. But there’s no way to know if new methods will appear that are even more effective and can easily get confidential information in detail.

Limiting sensitive data use in AI is a much safer approach. And the amount to limit it depends on what kind of account you're working with and what agreements are in place.

 

Sensitive data in different AI account tiers

Different AI providers and models handle data differently, so there's no definitive way to determine the risk based on account tier alone. However, as you'd expect, the higher-level accounts offer more security options to help manage and protect how data is used. 

 

Free accounts

As expected, free tools are the worst offenders, because nothing is truly free. Yet, companies like OpenAI can still claim they don't train on customer data, even on the free tiers, which might make it seem like there's minimal risk.

There's a bit of sneakiness involved here though. For instance, in the case of OpenAI, it's possible to not have your data be used to train models on the free tier, but it's an explicit opt-out setting you have to turn on. So unless it's turned off, that data is at risk, and even still, it's not risk-free. 

Some don't offer an opt-out option at all. MetaAI, for example, writes that they train their model off public Facebook and Instagram posts and can include PII and other sensitive data.

It's not always obvious how user data is used and retained with free models, and you'll often have to go hunting through documentation and or terms and conditions to find out for certain. For simple,nonsensitive queries like asking who starred in a movie or when an event is, there's not much risk, but start putting in confidential information like health records and proprietary company details into a free account and you risk it being leaked. 

Even when the data should be safeguarded, there's always the risk that gets exposed. Some Claude chats have ended up on Google as recently as July 2026. In this case, chats not marked as “keep private” became publicly searchable for a time, and some of these chats included private research and patient health records. A similar situation previously happened in 2025 with ChatGPT. For a time, sensitive data was not only exposed, but easily searchable. 

The safest way is to avoid using sensitive data with free tools at all. 

 

Business and enterprise accounts

Business-tier accounts can be particularly dangerous for sensitive data because paying for a higher business plan. But a higher-tier plan doesn't necessarily mean you can send the tool anything and it won't be exposed. What's important is how the data is processed, and what guarantees you have.

This is where a DPA (data processing agreement) is extremely helpful. A DPA outlines how the processor uses, stores, and processes your data, and is a legally-binding contract. When there isn’t a DPA in place, there’s no guarantee that data won’t be exposed. The DPA protects your organization and the data and gives you legal recourse if broken.  

While DPAs provide a considerable amount of protection for your data, it's a still good idea not to give the AI tool more data than it needs. If you can accomplish your task with less sensitive data, it's the safer course of action, though of course AI gives better results when it can access company datasets. 

Enterprise tiers provide the most protection, and you can have much more confident that sensitive data isn't leaking. However, this doesn't necessarily solve the monitoring problem to see how users are interacting with AI tools, so you can still have visibility gaps. But Enterprise-tier data security can range from good to extreme Zero Data Retention for the most advanced protection. 

 

What are Zero Data Retention (ZDR) agreements? 

All of the major AI companies offer what they call "Zero Data Retention", which is far more legitimate than a vague promise. Like the name suggests, no data is retained when the setting is enabled. For managing sensitive data in AI, there probably isn't a better option, but keep in mind that it comes at a price; and not just in money. 

The biggest cost is that for ZDR to work, a lot of features stop functioning. Many features, like referencing stored files and continuing chats based on historical data, rely on having at least some data stored, so some of these features are limited or disabled as a tradeoff. 

Instead, you can get more out of your AI tools by controlling how sensitive data can be used in tools. 

 

Controlling sensitive data use in AI

Instead of solely relying on vendors, there are a few things you can do to manage sensitive data use in AI. We cover these in more detail in our article about monitoring and managing Shadow AI. 

The simplest is to ban all AI tools that you don't have a DPA with. If users can't use the tool, they can't send sensitive data to it. Unfortunately, this isn't always as easy it sounds and determined users can find workarounds, and often ones you'll have no idea about. 

Another is to review your data retention policies. This is particularly helpful when managing file uploads to AI. By default, files will be stored by the AI provider for a certain amount of time. You can control this period to fit your security profile. Prompts themselves are more difficult to restrain, especially if you don't have any visibility into what's being sent and received. 

Education and company policies can help, with the added benefit of helping employees see the why behind the action. However, to be most effective these need to be refreshed regularly to remind the team about the risks of using sensitive data with AI. Sure, it's helpful, and you can try and control how sensitive data is used in AI by implementing policies and procedures. 

Or you can go the other way and limit the AI from using sensitive data with a tool like Oveka. 

 

Use Oveka to stop sensitive data from going to AI

While you can ban and limit AI tools except for company-approved ones with the right policies in place, Oveka gives you several other options as well. Yes, you can restrict users from using non-sanctioned AI with Oveka. But you don't have to. 

Some other options with Oveka include: 

  • Limit the content types that can go to AI
  • Block file uploads
  • Stop AI from accessing specific datasets
  • Control AI access by user and group membership
  • Redact sensitive data that would've been exposed

With Oveka, you have full control over how you manage sensitive data with your organization's AI tools. If you're confident in your setup, you could even allow users to continue using some of the free tools they've grown accustomed to and add guardrails around usage so they can use them safely and stop sensitive data going to AI. 

For help with this or anything else with Oveka, you can get in touch with our team and we'll be happy to walk you through how to use Oveka to enable your team to use AI safely. Reach out to us or book a demo today