Shadow AI Is Already on Your Endpoints

Taylor Karl / Friday, May 22, 2026

/ Categories: Resources, Artificial Intelligence (AI)

16 0

Key Takeaways

Shadow AI Ownership: Endpoint teams own the control layer closest to where unauthorized AI tools run
Infrastructure Already Exists: Most organizations can detect and govern AI usage with tools they already have
Shadow AI Has Predictable Entry Points: Browser extensions, desktop apps, and locally run models are the most common entry points
Policy Backs the Controls: Endpoint controls without documented governance erode over time
Metrics Prove the Work: Detection time, approved tool usage, and endpoint coverage make governance measurable

A security ticket came in on Tuesday morning. Someone in Finance had been using a generative AI tool to summarize internal reports for months. The tool wasn’t on any approved list. IT had never seen it. The employee hadn’t asked for permission.

The endpoint manager pulled up the device. Full management coverage. Every configuration met requirements. They had complete visibility into the machine, but no policy governing unapproved generative AI use.

The gap exists because no one made AI governance a priority for the endpoint team.

Shadow AI is spreading across organizations the same way Shadow IT did a decade ago, but the stakes are higher. Employees adopt AI tools that fit how they already work, often before IT knows they exist. Those tools spread department by department, and IT ends up managing problems when something goes wrong.

Finance, Marketing, HR, and Operations are all running AI tools on devices IT already manages. The data those tools are sending outside the organization is the problem many endpoint teams haven’t addressed yet.

For endpoint managers, the infrastructure to address this already exists. Platforms like Microsoft Intune, Configuration Manager, and Defender for Endpoint already have the capabilities to detect and govern AI activity. The gap is in how they’re configured, not what they can do.

Why Shadow AI Becomes the Endpoint Team’s Problem

Shadow AI follows the same structural pattern as Shadow IT, but the exposure isn’t just larger. It’s fundamentally different. With Shadow IT, the risk centers on the unauthorized tool. With Shadow AI, the risk is what employees feed into it: data that leaves the organization's security perimeter and, in many cases, can't be retrieved once it's gone.

Employees adopt AI tools because they work and don't wait for approval. When something goes wrong, accountability follows a predictable path back to those with device-level access and the tools to find out what happened: the endpoint team.

Typically, endpoint teams don’t think of themselves as governance owners, but they should. The devices they manage are the exact point where Shadow AI enters, operates, and exposes data. The governance role and the authority to act come with the job. No policy change is required to start.

The risk an unauthorized tool carries depends on how it spreads and what it exposes. An unauthorized project management tool primarily captures workflow data and is easier to contain once discovered. Generative AI tools spread across an organization faster than most IT teams can track.

Device-level oversight is only useful if the team knows what to look for and where to find it. Without that visibility, unauthorized tools run undetected, data leaves the organization, and the endpoint team has no record of what happened or when. Finding and closing that gap is where the endpoint team's governance role begins.

Why Shadow AI Becomes the Endpoint

Map the Visibility Gap

AI-specific detection is where many endpoint environments fall short. Without broadening the scope of their controls, endpoint teams have no insight into which tools are running, how often, or which data they touch. Unmanaged devices and BYOD add another layer of complexity, and organizations with significant BYOD usage need to treat them as a separate workstream.

Endpoint environments already have detection mechanisms in place, but they’re rarely pointed at AI-specific activity by default. Knowing where Shadow AI consistently appears, and which tools already cover those entry points, is where the mapping starts.

Shadow AI activity consistently appears across five entry points:

Browser extensions: Extension inventories and browser management policies
Web-based AI tools: DNS and proxy logs via SIEM or proxy integration
Desktop applications: Software inventory and execution event logs
Locally-run models: GPU usage spikes, large model downloads, and EDR process activity
API wrappers on developer machines: Environment variables, outbound connections, and EDR network behavior

Mapping these entry points shows the endpoint team where coverage already exists and where the gaps are.

Browser extensions and web-based tools are some of the highest-volume entry points and the right place to start. Locally-run models and API wrappers on developer machines are higher risk and, in engineering-heavy environments, becoming common. They often handle more sensitive data and are harder to detect through standard inventory tools.

Even with existing capabilities, if no one makes AI-specific detection a priority, gaps emerge. The tools to close those gaps aren’t new. They’re already running on managed devices, waiting to be pointed at the right problem.

The Controls Endpoint Teams Already Have

Application control, policy enforcement, and device monitoring are the three controls that cover Shadow AI activity. Endpoint environments that already run these tools have what they need. Closing the Shadow AI gap requires deliberate configuration, not new tooling.

The first instinct when Shadow AI shows up on managed devices is to block it, but acting before building end-user buy-in can create bigger problems down the road. Employees who feel blindsided by restrictions look for ways around them. The monitoring phase gives endpoint teams time to research unapproved tools, understand why employees are using them, and address gaps in the approved stack before enforcement begins.

Each control targets a specific layer of Shadow AI activity:

Application control: Blocks or flags unauthorized AI tools at install or execution via App Control for Business (formerly Windows Defender Application Control/WDAC)
Policy enforcement: Governs clipboard behavior, upload controls, browser extension allowlists, and network filtering through Intune and Configuration Manager
Device monitoring and alerting: Flags outbound connections to AI endpoints, large data transfers, and repeated interactions with unapproved AI domains via EDR and SIEM

Combined, these three mechanisms give the endpoint team direct oversight of where Shadow AI is running and the resources it’s accessing.

The three controls cover the execution layer, but data classification sits outside it. DLP and sensitivity labels intersect directly with endpoint controls, and teams should coordinate with whoever manages them. Controls and classification only work together when a documented policy connects them. Without it, exceptions accumulate and governance erodes.

Back Your Controls With Policy

Technical controls without a policy behind them are easy to work around. Employees who want to use an unapproved tool will find a way if there’s nothing formal preventing it. A documented AI usage policy gives the endpoint team the authority to act and a clear standard to enforce against.

Shadow AI governance follows a clear chain, with each function linking to a specific role. Security defines risk thresholds, Legal and Compliance define the boundaries for data handling, and business units identify approved use cases and the tools that support them. When each function owns its piece, the endpoint team has something concrete to enforce.

When those functions operate in silos, endpoint teams end up enforcing policies that don’t reflect real risk thresholds or approved use cases. The role only works if the policy behind it is operational and specific enough to act on.

An enforceable AI usage policy covers five areas:

Approved tools and licensing tiers: A maintained list of permitted tools, the conditions for each, and which data categories they may touch
Prohibited data categories: Classification-based restrictions tied to sensitivity labels
Device compliance prerequisites: The baseline a device must meet before the endpoint team grants AI tool access
Exception request workflow: How employees request access to unapproved tools and what the approval path looks like
Review cadence: A defined schedule for evaluating the approved list as AI tool categories change

Together, these components give the endpoint team a clear mandate and the structure to act on it.

A policy without measurement is an assumption. Endpoint teams that can’t show whether their controls are working can’t defend their governance decisions when it matters most. Defining what the team controls and proving it works is what makes governance defensible.

Measure What You’ve Closed

Knowing a policy exists isn’t the same as knowing it works. Metrics give endpoint teams the evidence to show whether controls are holding, whether approved tool adoption is growing, and whether enforcement is keeping pace with Shadow AI activity. This evidence is what turns governance into something the team can stand behind.

The metrics worth tracking are those the endpoint team can influence. Broad security metrics tell an organizational story, but they don’t show the specific gaps endpoint teams detected and closed. The indicators below give teams a concrete way to show what changed and why it matters.

Four indicators show whether governance is working:

Mean time to detection and response: How quickly the team moves from spotting a new AI tool to enforcing policy against it
Approved tool usage rate: The percentage of AI activity occurring on sanctioned tools, the clearest indicator that governance is shifting behavior
Exception request volume: Fewer requests over time signal that the approved tool list meets employee needs
Endpoint coverage: The percentage of managed devices running AI-specific controls, which shows whether enforcement is complete or still patchy

Taken together, they trace the team’s work from detection through enforcement.

Metrics are only useful if someone acts on them. Reviews of both metrics and policy should run on the same schedule. If the data shows enforcement gaps or exception request volume increasing, revisit the approved tool list now, not at the next calendar quarter.

Showing what was running three months ago, what’s running now, and where things are heading makes governance work visible. That’s how endpoint teams demonstrate they don’t just manage devices. They control organizational risk.

Build the Capability Once

Endpoint management has a maintenance reputation, and Shadow AI is changing it. Routine tasks like patching, configuration, and compliance maintenance have led leadership to underestimate how much the role controls. The infrastructure endpoint teams manage sits at the exact point where organizational risk gets controlled or exposed. Teams that make Shadow AI governance a priority own that layer.

New Horizons partners with IT and endpoint teams to build the technical skills Shadow AI governance demands. Governance is a skills challenge, and teams that invest in building those skills are the ones that move from managing endpoints to governing them. If you need your team to close this gap, we can help prepare them for it.

Ready to turn your endpoint team into your organization’s first line of defense against Shadow AI?

New Horizons gives endpoint teams the skills to turn existing tools into a complete Shadow AI governance capability.