The Shadow AI Problem: Finding What's Already Running
Shadow AI lives in config files, extensions, OAuth grants, and local models that never cross the network. Here's how to find it and govern it.
The Neo Team
Key Takeaways:
- Most shadow AI is started by a person at a device and never crosses the network, so discovery starts on the endpoint, where MCP configs, skills folders, and local model ports give it away.
- OAuth grants, model-provider API keys, and the admin consoles of sanctioned SaaS platforms each hold their own inventory.
- An inventory entry is only useful once it records who runs the tool, what it can reach, what it can do, and where its data goes.
In Verizon's 2026 Data Breach Investigations Report, 45% of employees were regular users of AI on their corporate devices, authorized or not, up from 15% a year earlier. Of those users accessing AI services, 67% did so through non-corporate accounts on corporate devices. Shadow AI ranked as the third most common non-malicious insider action in the report's data loss prevention dataset, a fourfold increase on the prior year.
Part of the gap is that nobody has said what approved means. In Stack Overflow's 2026 developer survey of more than 30,000 technologists, only 24% said their organization publishes approved tools or an official AI policy or guidance.
Even organizations required to keep an AI inventory miss what's already in use. In March 2026, the US Government Accountability Office found that contracted AI tools used in the Internal Revenue Service's criminal cases were absent from its inventory of 126 use cases. None of the bodies overseeing its AI use cases was responsible for managing AI investments across the agency.
Know what counts as shadow AI now
Custom agents, Model Context Protocol (MCP) servers, skills, coding agents, browser extensions, and local models now sit alongside the chat apps, and most of them were started by a person at a device rather than by another system. The table shows types of shadow AI and the layer that sees each one.
| Type | Where it lives | What finds it | What misses it |
|---|---|---|---|
| AI apps used through personal accounts | The browser, on a corporate device | Proxy and secure web gateway logs, data loss prevention (DLP) tools | Identity provider logs, since the account isn't yours |
| AI features inside sanctioned software as a service (SaaS) | The vendor's admin console, often enabled by default | The vendor's own console | Network tools, since the traffic is already allowed |
| Browser extensions | The browser profile | Browser management, endpoint scans | Proxies, which see the sites, not the extension |
| Coding agents and CLIs | Package managers, developer home directories | Endpoint inventory | Network tools, which see a call to a model's application programming interface (API) but not which agent made it or what it was given |
| MCP servers and skills | JSON config files and skills folders per agent | A file scan of known config paths | Network tools, since local servers run over stdio and skills are just files |
| Local models | A model runner bound to localhost | Endpoint process and port inventory | All network tools |
To find shadow AI, start on the endpoint
The endpoint holds the richest and least used evidence. Every major coding agent writes its MCP servers to a known file. The MCP project's own documentation walks through one desktop client and places its servers in a JavaScript Object Notation (JSON) file in the user's application data folder.
Skills follow the same pattern, with a cross-client convention of .agents/skills/ at user and project level. Local models are just as findable, since a model runner such as Ollama binds to port 11434 on localhost by default.
Local MCP servers run with the user's own permissions, which means a server listed in a config file can do anything that the user can do. Integrated development environments (IDEs) can also import each other's configurations. In VS Code, an autodiscovery setting can read the config files of other clients, including that desktop client. One server can therefore seed several agents from a single file.
A device-management query can collect all of this in one pass, helping to shape a usable endpoint inventory.
Find shadow AI through OAuth grants and API keys
Where the endpoint shows what is installed, the identity layer shows what has been granted through Open Authorization (OAuth), the standard most AI apps use to connect to a workspace. An OAuth risk report by security vendor Material Security covered 22,332 applications connected to 21 enterprise Google Workspace tenants.
Of those, 47.2% had recorded no active usage in 90 days and remained connected. More than a thousand still held live tokens for users who no longer existed, because account suspension doesn't revoke a third-party OAuth grant.
Popular MCP setup guides often tell developers to put API keys directly in the config file. In secrets-detection vendor GitGuardian's 2026 secrets sprawl report, 24,008 unique secrets sat in MCP-related configuration files on public GitHub, 2,117 of them valid.
Pull the OAuth grants from your identity provider and list the model-provider API keys in your secrets manager and in developer config files. Together they form a second inventory that the endpoint scan can be checked against.
Check the consoles you already own
Sanctioned platforms now frequently ship agents of their own. In Microsoft's 2026 Work Trend Index, active agents in its 365 ecosystem grew fifteenfold in a year. In May 2026, Google released an AI control center for Workspace that shows who is using AI across its own apps. Each console sees AI inside its own estate and nothing on the endpoint.
The admin pages of your top SaaS vendors are therefore a third inventory source, and each AI feature they switch on should trigger the same review a new vendor would.
Map what each tool can do and reach
For every agent, server, extension, and model on the list, record the following:
- Who runs it. The human account the agent, server, or extension runs under, and the person behind that account if it's a service identity.
- What it can reach. The files, repositories, and SaaS data it can open through the user's permissions, OAuth scopes, and the keys in its config.
- What it can do. Whether it only reads, or can also write files, run shell commands, send messages, or call other tools without approval.
- Where its data goes. The model providers and outside services it sends prompts and outputs to.
An Ontario government audit published in May 2026 shows why each entry needs this detail. Between April and August 2025, staff accessed about 400 AI websites, and about 60% of those sites were rated unsafe or unsecured. Only 6% of their generative AI website use went through the approved tool.
How Neo finds and governs shadow AI on every endpoint
Neo's discovery loop answers the first question, what software is running, on every endpoint. Neo Node inventories the agents, command-line tools, IDE assistants, extensions, MCP servers, skills, and local models on each device, including the ones that never touch the network. It checks each against Neoverse’s catalog of 4M+ agentic artifacts, each with a risk profile.
Most teams reach a full inventory in under 5 minutes from the first scan. Each entry carries the user who runs it and the data it can reach, which answers who runs it on day one.
The same Neo Node then governs what each tool may do. Posture checks flag over-broad permissions and auto-approve modes, and the policy allows, asks, or blocks each action at the tool call, before it executes. Every decision is recorded with the human, agent, and action behind it.
Get a demo to see what's already running on your endpoints.
FAQs
What is shadow AI?
Shadow AI is any AI application, model, agent, extension, or integration used for work without the organization's knowledge or approval. It includes personal accounts on public AI services, AI features switched on inside sanctioned software, and browser extensions. It also includes coding agents, the MCP servers and skills they load, and models running locally on employee devices. Much of it produces no network traffic, which is why it goes unseen by tools built to find shadow IT.
Should shadow AI be blocked once it's found?
Not by default. Most of what an inventory turns up is people doing their jobs with a tool that worked better than the approved one, and a block with no alternative pushes the same work onto personal agents, where no control reaches it. With Neo, the tool can stay in use, and only the actions flagged by policies are held for approval or blocked.
How often should an AI inventory be refreshed?
Continuously where possible, and on triggers where it isn't. A new agent, extension, or MCP server can appear on a laptop in an afternoon, and an OAuth permission outlives the employee who created it, so a quarterly review leaves months of exposure. The Neo Node tracks AI tools on each endpoint in real time, so a new agent or MCP server is detected without waiting for the next scan.