Big challenges loom for business leaders as artificial intelligence continues to be incorporated into organizations seeking to improve efficiency. Governance. As companies discuss frameworks and evaluate vendor solutions, employees are already making their own decisions.
our recent research Almost half (49%) of employees use AI tools at work that are not authorized by their employer. And most of the time their motives are clear. They want to do their jobs more efficiently and effectively. However, the rise of this hidden new threat vector represents one of the most significant security gaps facing organizations today.
The new face of shadow IT
Security teams have been working on shadow IT for years. When employees find company systems too cumbersome, they download apps, share files through unauthorized services, or find workarounds. In fact, our research found that 63% of respondents believe it is acceptable to use AI tools without IT oversight if the company does not offer a sanctioned option.
Since data is typically confined within predefined boundaries, these risks can be significantly reduced. Shared documents and unauthorized messaging apps posed compliance issues, but the potential damage was limited.
LLM has rewritten the rules. Employees paste their own code into something like ChatGPT, upload customer data to analyze trends, and share strategic plans with LLMs to refine their writing. Doing so circumvents IT policies and exposes an organization’s most valuable assets to systems that can hold, learn from, and regurgitate that information.
This is such a widespread issue, in part, because employees believe they are contributing to the company by using AI tools. Our research also found that 71% of employees believe the efficiency benefits of using unapproved AI tools outweigh the privacy risks. They’re trying to work smarter, but the problem is that their efficiency efforts can put the business at risk in ways they didn’t foresee or consider.
Problems hidden in plain sight
The limits of employee perception when it comes to handling AI data should be a concern to any security leader. Just over half (53%) of employees understand how the data they input into AI tools is stored, analyzed, and stored. This points to a clear knowledge gap that organizations are failing to address.
Let’s consider what this means in practice. Employees upload spreadsheets containing customer information to an AI tool to help create presentations. Do they know if that data is part of the tool’s training set? Can it be reconstructed from the model? Will it also appear in responses to other users?For most employees, these questions will not be seen as a concern.
LLM introduces a new category of insider risk that traditional security controls were not designed to address. Unlike traditional applications, where data flows can be monitored and controlled at the network level, interacting with AI tools can completely bypass corporate security. An employee working in a coffee shop using a personal laptop can potentially reveal trade secrets without triggering a single alert.
Information does not pass through these systems and disappear. Prompt history may be recorded, and even if the provider claims not to use customer data for training, the data still exists outside of corporate governance. Adding to the complexities of data security, you may be subject to changing terms of service and legal jurisdictions that require disclosure.
Compliance dimensions
For regulated industries, the compliance implications are particularly significant. Financial services companies handling customer data under the GDPR, healthcare organizations bound by data protection requirements, and businesses managing sensitive information face serious risks when employees transfer that data through unvetted AI systems.
The problem is already a reality. We have seen cases where AI tools inadvertently leak sensitive information through responses to other users. security researcher We demonstrated that training data can be extracted from language models using appropriate prompting techniques. If that training data contains sensitive information, the implications extend far beyond awkward conversations with regulators.
Beyond detection to prevention
Traditional approaches to shadow IT rely heavily on detection, such as finding rogue applications and blocking access. This has never worked perfectly, and is even less suited to the realities of the AI era. Browser-based AI tools are increasingly indistinguishable from legitimate web traffic, and employees can access them from personal devices using personal accounts, on networks not controlled by the organization.
Detection alone is meaningless. Once an employee is found to have used a fraudulent AI tool, the damage has already been done. Sensitive data is no longer in the environment and the information entered into the LLM context window cannot be recalled.
The answer lies in prevention. Data leakage prevention (ADX) technology can identify sensitive data at the endpoint before it leaves the device, regardless of which applications employees are using. Real-time detection, automatic policy enforcement, and blocking unauthorized data movement without business interruption are key requirements.
The human element is also important. Especially since most employees don’t truly understand the security implications of using AI. So instead of expecting every employee to be a security expert, organizations need systems that automatically provide guardrails.
Governance challenges
Beyond technology, organizations need robust policies that recognize the realities of using AI while preventing data loss. This starts with allowing employees to use AI tools as they see value, rather than adopting blanket bans that would be ineffective and counterproductive. The goal is to use AI responsibly with appropriate guardrails in place.
In fact, companies must provide approved alternatives that meet true productivity needs and maintain security controls. If your employees are using ChatGPT to improve their writing, provide an enterprise AI solution that ensures proper data processing. If you are using non-approved tools to analyze your data, we provide approved options that do not expose sensitive information. At the same time, organizations must deploy AI security tools that can monitor and control this access. Trust, but verify.
Education is equally important because if employees don’t understand the risks, they can’t be expected to make informed decisions about using AI. Training programs should go beyond general warnings about data security and provide concrete examples of what can go wrong when sensitive information enters an AI system.
This requires being honest about the threat situation. Conversations should focus on protecting the organization from risks that it does not fully understand. When staff understand that increased efficiency can lead to data breaches, damaging customer relationships and triggering regulatory fines, mindsets change.
The gap between what employees are doing and what security teams know is widening and will only widen if organizations don’t act quickly. Shadow AI requires immediate responses, from endpoint protection and clear policies to education programs that help employees understand the risks.
