67% of Your Employees Are Using AI You Know Nothing About

You probably have an AI policy. Maybe you rolled it out last year. Maybe it covers ChatGPT, maybe it mentions data classification, maybe it has a section about not sharing customer information with external tools.

Here is the uncomfortable question: do you have any idea whether people are actually following it?

The 2026 Verizon Data Breach Investigations Report (DBIR) looked at this closely. What it found should make any GRC or security leader put down their coffee and read carefully.

In 2025, the year covered by the latest DBIR, 67% of employees were using personal accounts to access AI services on corporate devices. Not company-approved tools, not managed accounts your IT team can see. Personal accounts. Their own ChatGPT account, their own Gemini login, their own subscription to whatever AI tool they discovered last week. The company has no visibility into what goes in, no visibility into what comes out, and no record of any of it.

That is not a rogue employee problem. At 67%, that is a cultural norm.

This Is Not a New Problem. It Is a Much Bigger One.

Shadow IT has been around for decades. People download software their IT department hasn't approved. They use personal Dropbox accounts to share files. They find workarounds because the approved tools are slow or clunky or just not as good as what they can get on their own.

Shadow AI is the same workaround, but the scale of what can leak is completely different.

When someone used an unapproved file-sharing app in 2015, they might move a presentation or a spreadsheet. When someone used an unapproved AI tool in 2025, and that behavior surely continues into 2026 as we’ll see in next year’s DBIR, they might paste in an entire codebase to get help debugging it. They might upload an internal strategy document to get a summary. They might feed proprietary customer data into a prompt because they are trying to finish a report faster and the AI is genuinely helpful.

They are not trying to cause a breach. They are trying to do their jobs.

The DBIR makes this concrete. The most common type of data submitted to unauthorized AI tools in 2025 was source code. Source code. Not generic documents, not vague business content. The actual technical foundations of products companies have spent years building. In 3.2% of DLP events analyzed, researchers found internal research and technical documentation being uploaded to AI systems that the company had no relationship with and no contractual protections from.

That is intellectual property walking out the door through an AI chat, and most organizations have no idea it is happening.

The Browser Is Now a Risk Surface Too

Here is the part that tends to surprise people even more than the account-sharing statistics.

The DBIR found that in 2025, the average company had more than 15% of its users running unauthorized AI extensions in their browsers. Browser plugins. The kind of tool that sits quietly in the corner of your screen and watches everything.

Many of these extensions are built to collect context. They read what you are browsing so they can give you better suggestions. They remember what sites you visit so they can be more helpful next time. That is how they work, and in a personal context it might be a reasonable trade-off.

On a corporate device, browsing internal systems, that same functionality means an AI plugin you never approved may be quietly ingesting data from your internal portals, your project management tools, your HR systems, or wherever else your employees spend their day. Not maliciously. Just by design.

Most security tools are not looking for this. Most DLP configurations were written before AI browser extensions existed as a category. The visibility gap is real, and it is wide.

Shadow AI Grew Fourfold in a Single Year

If the 67% number feels alarming on its own, consider thetrajectory.

In 2025, Shadow AI became the third most common non-malicious insider action detected in DLP datasets. That is a fourfold increase compared to the previous year. And the share of employees using AI regularly on corporate devices, whether the tools are approved or not, jumped from 15% to 45% in a single year.

That pace of growth is not something a policy written once and published on an intranet can keep up with. The behavior is spreading faster than most governance programs can respond to it.

This is what makes Shadow AI genuinely different from most compliance challenges. Most risks grow gradually. You update your controls, you run awareness training, you close gaps incrementally. Shadow AI is moving at a speed that makes incremental responses look like standing still.

Why Your AI Policy Is Probably Not Enough

Most AI policies share a common structure. They define approved tools. They say what data classifications can be shared externally. They remind employees not to input confidential information into public AIsystems.

That is a reasonable start. It is also never sufficient.

The policy exists as a document. What it usually lacks is any mechanism to know whether actual users’ behavior matches the document. There is no inventory of which AI tools are actually in use across the organization. There is no monitoring of what data types are being submitted to external AI systems. There is no way to detect when a new AI browser extension shows up on a managed device and starts doing what it was designed to do.

A policy without visibility is not governance. It is a hope.

Real AI governance means knowing what AI tools exist in your environment, including the ones nobody approved. It means being able to see when sensitive data is submitted to an external AI system, and having the ability to act on that in something close to real time. It means treating AI usage as a continuous monitoring problem, not a policy publication problem.

The regulatory environment is also pushing in this direction. The EU AI Act creates obligations around high-risk AI usage that require documentation and control evidence, not just stated intentions. GDPR's data processor provisions apply to third-party AI tools that receive personal data, regardless of whether that sharing was intentional or accidental. Organizations that cannot demonstrate actual control over how their employees use AI tools will face an increasingly difficult conversation with regulators.

What Closing the Gap Actually Looks Like

The starting point is honest assessment. Most organizations genuinely do not know which AI tools are running in their environment. Getting that inventory, including browser extensions, standalone applications, and web-based services accessed through personal accounts, is the first step. You cannot govern what you cannot see.

From there, the work is about building controls that match the reality of how people work. Blanket bans on AI tools tend to backfire. People find workarounds, and now you have the same behavior plus the organizational friction that comes from fighting a tool people find genuinely useful. The more durable approach is building a governed path. Make it easy to use approved tools. Make those tools good enough that the unauthorized ones are less appealing. And monitor the edges for behavior that falls outside the policy.

The DBIR closes its Shadow AI section with a line worth sitting with. It notes that just because there is a new and exciting tool does not mean organizations should abandon decades of data governance and third-party risk management practice. The principles are not new. The application of them to AI just requires building new plumbing.

That plumbing is not optional anymore.

--------------------

‍Niraj Mehta serves as Managing Director for EMENA and APAC at LockThreat. Previously, he founded ThreatReaper, an AI security and governance platform that was acquired by LockThreat. This makes Niraj a serial entrepreneur and 2x founder with two successful exits. Earlier in his career, Niraj served as a co-founder at Ozone, where he focused on enterprise DevSecOps, cloud security, and application security solutions for large organizations. Over the years, he has advised enterprises across EMEA on cloud security, AI governance, cybersecurity, and regulatory compliance. His unique combination of entrepreneurial leadership, deep technical expertise, and enterprise consulting experience provides a practitioner-led perspective on AI security, governance, and the complex compliance challenges organizations face as they scale AI adoption.

--------------------

Data in this post is drawn from the 2026 Verizon Data Breach Investigations Report. Shadow AI statistics are from pages 13 and 60. DLP findings, including source code and research document uploads, are from pages 60 and 61.