Your Employees Are Feeding Company Secrets to Unlicensed AI Models

It's an everyday occurrence: employees who have no idea what happens to the company data they upload to various AI services. It's a widespread problem, but what can company leadership do about it?

Share
Your Employees Are Feeding Company Secrets to Unlicensed AI Models

Legal documents, HR data, source code, and other sensitive company information are being fed into unlicensed, publicly available AI models at a rapid pace, creating a growing headache for IT managers and company leadership. A new study shows that employees at many organizations are using unauthorized AI models on their own private accounts in the workplace.

Risks of Unlicensed AI Models

Employees are sharing company legal documents, source code, and employee information with unlicensed AI tools like ChatGPT and Google Gemini, which can cause major problems for IT and company leadership, according to research from Cyberhaven Labs. About 74 percent of ChatGPT use at work happens via private accounts, potentially making this data available for AI training. The corresponding figure for Google's Gemini and Bard is over 94 percent.

Growing Use of AI Tools

Nearly 83 percent of all legal documents shared with AI tools go through private accounts, and roughly half of all source code, R&D material, and HR and employee records go to unauthorized AI tools. The amount of data fed into AI tools increased almost fivefold between March 2023 and March 2024.

Where Does the Data Go?

Many users are unaware of what happens to their company data when they share it with unlicensed AI. ChatGPT's terms of service state that users own the content they input, but the AI may use that content to improve its services, meaning it may train on this data. Users can, however, opt out of this AI training, though far from everyone knows how. And can you even trust the big companies about what actually happens to the data?

Security and Privacy Risks

There are few rules governing what AI developers can do with users' data, which worries security experts. The coming wave of AI developers may be less trustworthy and have weaker cybersecurity protections than major players like OpenAI and Google. Brian Vecci, CTO at Varonis, points out that future AI tools could carry higher risks for data handling.

Risky Behavior and the Need for Control

Sharing company or customer data with unlicensed AI companies creates risks, regardless of whether the AI model trains on that data or shares it with other users. Pranava Adduri, CEO of Bedrock Security, recommends that organizations sign licensing agreements with AI vendors to limit data use and enable safe handling of the data.

Training and Security

One of the fundamental challenges is that employees simply don't know the risks. Good training on what constitutes sensitive data, how it risks being used, and how employees should handle it is a big step in the right direction. It can be general but needs to be tailored to each company's circumstances. The Swedish Tax Agency, Volvo, and a small online retailer have completely different data-handling requirements.

So How Do You Best Protect Yourself?

AI-certifying both company leadership and employees so they have solid knowledge of data handling and tools is the first step. Taking inventory of data handling and implementing access controls, as well as running, for example, open-source models on local servers, is another step in the right direction toward creating a sound environment where AI can be used safely.