Author: Julia Grits, Netwrix Brand Manager
ChatGPT suggested the title, so I don’t think it will mind. But let’s get one thing straight: no tool is to blame for data leaks. It’s just that users constantly spraying data everywhere overuse them without any restraint.
Shadow IT is dead. Long live Shadow AI.
Let’s start from the beginning.
Shadow IT arose long before the advent of generative artificial intelligence.
In the 2010s, employees began using their personal accounts for work on a massive scale: personal Dropbox and Google Drive accounts, private GitHub repositories, messaging apps, and file-sharing services. The reason was quite simple – most of the corporate tools were less convenient and fast. As a result, companies began to lose control over their files with where, how, and why they were stored.
With the generative AI, the problem has reached a whole new level. Whereas employees used to simply move files around, today they are feeding source code, technical documentation, event logs, business requirements, and other valuable information into ChatGPT, Claude, Gemini, and other AI services. Thus, Shadow IT has evolved into Shadow AI – a brand-new phenomenon of losing control not only over files and tools, but also over the knowledge and data used in daily work.
In the past, perimeter security was sufficient, but now a different approach is needed, and it’s not just about total control. Statistics show that 45% of users find a way to bypass app blocking.
Why is the IT sector hit the hardest?
IT companies have found themselves at the center of the problem. Unlike most other industries, these companies have employees who:
- have sufficient technical knowledge
- have enormous interest in new technological innovations
Developers, DevOps, QA-experts, business analysts and architects constantly work with vast amounts of digital data. In addition, they are looking for ways to automate their routine tasks. That is why they were among the first to actively adopt generative artificial intelligence – for writing code, analyzing errors, creating documentation and reports, and finding technical answers.
The problem is that in order to get an answer (preferably without any hallucinations), the AI service needs to be provided with context. In practice, this means providing snippets of source code, SQL queries, architectural diagrams, technical documentation or even customer data. Routine developer`s workflow is a transfer of intellectual property outside a controlled environment for the CISO and the cybersecurity team.
The situation is complicated by the fact that modern AI tools are increasingly being integrated directly into the work environment. IDE`s AI assistants, browser extensions, chatbots, and specialized code-generation services allow users to interact with artificial intelligence in a way that is virtually imperceptible to them. In many cases, employees don’t even think about what exactly they’re sending when making requests.
In addition: IT specialist always knows better than anyone else!!
To avoid offending anyone, let’s call it “a high level of sophistication in choosing one’s own tools”. If something helps solve a problem faster, you should use it; getting approval from the information security department is “just bureaucracy.” That’s why Shadow AI in IT companies usually arises not because of deliberate sabotage, but precisely because of the desire to work more efficiently.
For IT companies, controlling data transfers to AI platforms is now just as important as protecting email, cloud services, or external storage devices!
The three most dangerous scenarios
Scenario #1: Debugging via ChatGPT
“Here’s a snippet of code and the error log. Find the problem.”
And boom! the company’s code has left the controlled environment.
Scenario #2. Cursor and IDE’s AI
Modern IDEs regularly interact with external LLMs.
Developers don’t always know which files are being analyzed or where the data is being sent.
Scenario #3: Business Analysts and Project Managers
AI handles just about everything:
- contracts
- technical specifications
- financial calculations
- commercial proposals
I know from personal experience (guilty).
Shadow AI Control Formula
Let me throw in a few scarier real-life figures:
- 81% of employees may use unauthorized AI tools while on the job. And here’s the most interesting part: among security professionals, the figure reaches 88%!!
- 56% of organizations report that employees are uploading sensitive data to unauthorized SaaS services.
- The data shows roughly the same figures from LayerX: 45% of employees are already working with GenAI tools, and 77% of them regularly upload corporate data there.
Gartner predicts that by 2027, more than 40% of AI-related data breaches will be linked to the misuse of generative AI.
Against this backdrop, many companies are trying to tackle the issue of Shadow AI through restrictions: blocking specific AI services, limiting access to websites, or implementing AI usage policies. In practice, however, this approach rarely yields the desired results. The generative AI market is evolving too quickly: if an employee can’t use one service today, tomorrow they’ll find dozens of alternatives with similar functionality.
And here I could say that I’m in favor of education.
A whole bunch of cybersecurity training for employees, if we were talking about any other industry. But this is IT… see the “IT specialists always know better than anyone else.”.
And perhaps that is precisely the turning point. Perhaps it makes sense to acknowledge that you have a team of smart people who want to work more efficiently, not harder. Perhaps that is why a modern approach to security should focus not on controlling people or tools, but on controlling data and actions.
In simple terms, the risk of a data leak via Shadow AI can be described by the following formula:
Risk of a data breach = Sensitive data × Transmission channels × Lack of controls
The risk can only be reduced by addressing each of these components.
Step 1. Determine what data needs to be protected.
A company must understand where its source code, technical documentation, customer data, financial information, and other critical assets are located. Here’s a basic rule: you can’t protect what you haven’t identified.
Step 2. Monitor information transmission channels.
Today, data can leave an organization through more than just email or USB drives. Web-based AI services, browser extensions, IDE`s AI, cloud services, messaging apps, and even simple copy-paste operations have become new channels for potential data leaks. Understand the ways data can leave the organization.
Step 3. Implement ongoing monitoring and response policies.
The organization must be able to detect when a user attempts to transmit confidential information, assess the risk of such an action, and automatically apply appropriate restrictions – from a warning to blocking.
Conclusion
Modern DLP solutions will be built on these very principles. The goal is not to prohibit employees from using new technologies, but to ensure control over data flow regardless of which tool is used today or will be tomorrow.
In the context of Shadow AI, this approach will allow organizations to maintain a balance between productivity and data protection without stifling innovation through excessive control.
