Data-centric security for shadow AI: protect sensitive data wherever it goes
You can’t block employee access to every AI tool. You can make sure your sensitive data stays protected when it gets there.
Shadow AI reaches data you may not know exists, let alone know is sensitive. Knowing where your sensitive data lives and clearing out redundant, obsolete, and trivial (ROT) data gives your data security program a solid foundation. The next step is protecting the sensitive data you keep, even when copies end up in AI tools nobody approved.
This final post in our shadow AI series builds on what we covered about the rise of AI in the workplace and shadow data.
The sensitive data your business depends on, such as customer records, payment data, and health information, isn’t going anywhere. In fact, it’s multiplying. Data gets copied into analytics platforms, shared with partners, and pulled into AI workflows. Some of those copies will end up in places you didn’t plan for, outside of controlled repositories and networks.
That’s where data-centric security comes in. Instead of trying to control access to every AI tool, you protect the sensitive data itself, so it stays protected wherever it goes.
What is data-centric security?
Data-centric security is an approach that protects sensitive data directly, rather than relying only on the security of the networks, applications, or storage around it. Sensitive data can be encrypted, and tokenized, or sensitive values can be masked at a field level enabling protection stays with the data as it moves between systems. Authorized users and applications can access the original values when policy allows. Everywhere else, the data stays protected.
Perimeter and access controls still matter. They guard the places where data lives. The challenge with AI is how easily data can leave those places. Data-centric security covers that gap, because a protected copy carries its protection with it.
Why shadow AI makes this urgent now
Shadow AI changes the picture in two ways. First, the list of possible destinations keeps growing. New AI tools and features appear faster than any security team can review them, and employees who are trying to work faster will find them. Second, data moves quickly and casually: a paste into a chat window, an upload for a quick summary, an export to test an idea. Once data reaches an external tool, it can be difficult to get back. Recently, mathematicians have raised questions about whether their AI chat conversations may have contributed to work they didn’t author. And if you don’t know where all your sensitive data is, you can’t know what has already left.
Access controls such as blocking or allowlisting remain an essential part of any security program, but they can only act on the tools you know about. Shadow AI data security works best when you plan for the reality that some data will reach places you didn’t approve, and make sure the sensitive values in it are protected before it gets there.
How tokenization keeps data useful and protected
Traditional encryption has a practical drawback: it changes the format of data. An encrypted card number no longer looks like a card number, which can break the applications, databases, and analytics that expect one. That’s one reason sensitive data often stays in the clear.
Format-preserving protection addresses this. Tokenization replaces a sensitive value, such as a Social Security number, with a substitute that keeps the same length and format but reveals nothing useful on its own. Format-preserving encryption achieves a similar result using standards-based cryptography. Because protected values keep their structure, many applications can keep working without structural, system wide reworks, and consistent protection means records can still be joined and analyzed across data sets.
For AI data security, that has two benefits. Your teams can use protected data in analytics and many AI workflows without exposing the original values. And if a protected data set ends up in an unsanctioned AI tool, the tool receives tokens rather than the real customer details behind them.
Find, reduce, protect: data-centric security in practice
Data-centric security isn’t a standalone fix. It’s one part of a broader security approach, and it’s most effective alongside knowing what data you have and keeping only what you need:
- Find what’s sensitive. Discovery and classification show you where sensitive data lives, including the shadow data covered in our last post. OpenText™ Voltage™ Data Security Platform scans structured and unstructured repositories and classifies sensitive data, so you know where protection matters most.
- Reduce what you don’t need. Every record you retire is one less record to protect. OpenText™ Voltage™ Structured Data Manager helps you archive or retire inactive data from application databases under policy.
- Protect what remains. OpenText™ Voltage™ SecureData Servers applies format-preserving encryption, tokenization, and masking to sensitive data, with stateless key management across on-premises, cloud, and hybrid environments. Because protection is applied to the data itself, it can follow that data into analytics platforms, test environments, and AI pipelines while helping keep the workflows that depend on it running.
Why OpenText for data-centric security
OpenText brings discovery, data reduction, and data-centric protection together in one data security portfolio, so you can work through each step with a single partner instead of piecing together separate tools. Voltage format-preserving encryption is based on the NIST-standardized FF1 method, giving you standards-based cryptography that’s designed to protect sensitive data at enterprise scale.
Protect the data, and the tool matters less
Shadow AI isn’t going away, and neither is the pressure to adopt AI quickly. Across this series, the answer has stayed the same: focus on the data. Know what you have, retire what you don’t need, and protect what remains so it stays protected wherever it goes. That’s how you give your people room to use AI while keeping sensitive data under your control.
