Shadow AI is usually described as a people problem: an employee pasting company data into a chatbot nobody approved. That person exists, and the risk is growing: IBM’s Cost of a Data Breach Report 2026 found that incidents involving employee shadow AI more than doubled in a year, to 43% from 20%. But they are the more visible part of the problem.

The larger part is less obvious. The software your organization already runs has been adding AI features release by release, and each of those features sends your data to a model. They arrived the way every other feature arrives, in a vendor update, described in a release note that nobody in security read. That is the shadow AI this piece is about. A working definition: shadow AI is any AI model your data reaches that security has not inventoried, whether an employee brought it in or a vendor shipped it in an update.

You became an AI company without asking

Count the SaaS products your organization runs. Now count how many of them shipped an AI feature in the last twelve months. The CRM added an assistant. The helpdesk added summarization. The search box quietly became semantic. The analytics platform now answers questions in natural language.

Every one of those features means an LLM is being called with your data. Most of them arrived in a release note nobody in security read. Nothing was smuggled in. Nothing was unsanctioned. The procurement was approved years ago; the AI arrived later, by update.

And that is just the vendor side. Internally, the same quiet accumulation happens: a developer stands up an MCP server to test an integration and leaves it running. A team wires a model into a workflow as a proof of concept that becomes production by usage. A tool gets exposed to an agent for one project and never gets removed. None of this is malicious. All of it is invisible.

The natural-language tell: what shadow AI really looks like

Here is the field observation that changed how I think about this: a security engineer showed a customer their own application’s search feature (ordinary search, no chat window in sight) and pointed out that it was parsing natural language through a model. The customer’s response: “We’re AI-enabling our application?” They were. They just hadn’t filed it under AI.

That is the real shape of shadow AI. It does not look like a chatbot. It looks like search that got better, forms that autocomplete smarter, workflows that route themselves. If your definition of “AI application” requires a chat window, your inventory is wrong by a multiple.

Why shadow AI is riskier than other unmanaged assets

Unmanaged assets have always been risky. Unmanaged AI assets are risky in a specific, compounding way: each one is an uninspected path between your data and a model. Nobody is checking what flows in (customer records? credentials? regulated data?) or what flows out (into whose model, under what retention terms, trainable by whom?). An unmanaged web app is a door you forgot. An unmanaged AI integration is a door you forgot, that talks, and that can be talked into things.

This is why discovery is becoming the first-order requirement in AI security: before guardrails, before policy. The order of operations is unforgiving: you cannot apply a policy to an asset you have not found. You cannot govern what you cannot see.

Where to start: building an AI inventory to detect shadow AI

Start with a census, not a policy. Every place an LLM is called: customer-facing applications, internal assistants, developer tools, CI/CD pipelines, and the category most inventories miss: every SaaS product that added an AI feature since last year. For each: what model, whose infrastructure, what data in, what data out, which team owns it. For agents and AI tools running on developer machines, our open-source ShadowAI-Watch tool correlates local processes with outbound LLM connections, so the census can include what never went through procurement.

Expect the list to be roughly twice as long as anyone predicts. That gap between the AI you think you run and the AI you actually run is your shadow AI, and it was shipped to you, one release note at a time.

Thales’s Imperva AI Application Security is built around a simple ordering: discover every AI asset, assess the risk, then govern and protect. Our AI Discovery Playbook includes a self-audit worksheet for the census. Download it here.

The post Shadow AI Is Shipped to You, One Release Note at a Time appeared first on Blog.