Glossary
What is shadow AI?
Shadow AI is AI that employees use at work without security's approval. It includes coding agents, MCP servers, skills, local models and AI apps signed in with a personal account. Security cannot protect what it cannot see, and much shadow AI runs on the device, not on the network, so it never shows up in a gateway log.
By the AIBound team · Updated 18 November 2026
Key takeaways
- Shadow AI is AI that employees use at work without security's approval.
- Much shadow AI runs on the device, not on the network, so it never shows up in a gateway log.
- Each tool can reach company data, and nobody reviewed it.
Why shadow AI matters
Employees bring their own AI to work. The 2024 Work Trend Index from Microsoft and LinkedIn found that 78% of AI users bring their own AI tools to work. Each tool can reach company data, and nobody reviewed it. Shadow AI can run for months before an audit finds it, and it leaves no record for the auditor.
The risks of shadow AI fall into three groups. Data: an employee pastes customer records or source code into an AI app on a personal account, and the data leaves the company. Access: an agent connected to an MCP server can read files and systems far beyond the task it was given. Evidence: when an auditor asks which AI the company uses, there is no list to show.
Banning AI does not remove the risk. Employees who are blocked from one tool move to another, often on a personal account, where security sees even less. The safer path is to find what runs, decide on each one, and stop only the high-risk ones.
How shadow AI happens
Shadow AI is rarely a bad act. It starts when someone tries to get work done faster and the approved tools are slower or missing.
- 1 An employee installs an AI agent, skill or local model, or signs in to an AI app with a personal account.
- 2 The tool connects to files, code or company systems, often through an MCP server.
- 3 Security sees nothing, because the AI runs on the device and may never touch the network.
- 4 The tool keeps running. Nobody reviews it, updates it or removes it, and its reach grows each time the employee connects something new.
Shadow AI vs shadow IT
Shadow IT is any software or cloud service used without approval. Shadow AI is the same habit with AI, and AI can act on its own, which raises the stakes.
| Question | Shadow AI | Shadow IT |
|---|---|---|
| What is it? | Unapproved AI agents, models, skills and apps | Unapproved software and cloud services |
| What can it do? | Act on its own and reach files and systems | Store and move data a person puts in |
| Where does it hide? | On the device, often off the network | In accounts, invoices and the browser |
| How is it found? | By reading the device, where the AI runs | By reading network logs, invoices and sign-ins |
| What is the worst case? | An agent acts on company systems with nobody watching | Company data sits in a service nobody manages |
What are the risks of shadow AI?
Shadow AI carries four kinds of risk, and all four come from the same cause: the AI is there and nobody decided on it.
Data. An employee pastes customer records, source code or unreleased plans into an AI app on a personal account. The data leaves the company, and nothing records that it did.
Access. An AI agent borrows its owner's logins and files. Connect it to an MCP server and it can read and change systems far beyond the task it was given.
Supply chain. Skills and MCP servers are published and installed in minutes, often from public repositories that security never reviews. One bad skill can tell a trusted agent to read credentials.
Evidence. The EU AI Act, NIST AI RMF and ISO/IEC 42001 all expect a company to know which AI it uses and to show how it controls it. Shadow AI leaves nothing to show.
Time makes it worse. Reco's 2025 State of Shadow AI Report found that unsanctioned AI tools stay in use for more than 400 days on average.
What does shadow AI look like on a device?
On a device, shadow AI shows up as things that are installed or signed in: a desktop assistant, a coding agent in the terminal, an MCP server running locally, a skill folder that an agent loads, a local model such as Ollama, an IDE extension, a browser extension, or an AI app signed in with a personal account.
Each one has an owner, which is the person, a place, which is the device, and a reach, which is the files and systems it can touch. A useful finding answers three questions: what is it, who runs it, and what can it reach?
How do you detect shadow AI?
No single view finds all of it, because shadow AI lives in different places. Most teams combine these views and then add the one they are missing.
- 1 Read the device. Agents, MCP servers, skills, local models and IDE extensions run on the employee device, and much of it never touches the network. Reading the device is the only way to find these, and it shows who runs each one and what it can reach.
- 2 Read the network. Gateways and CASBs show traffic to AI websites and apps. They miss anything that does not send traffic through them, such as a local model.
- 3 Read accounts and spend. SaaS management and expense tools show the AI apps people paid for or signed in to. They miss free tools, local tools and anything on a personal account.
- 4 Ask employees. A survey or an amnesty week finds what people remember to mention. Treat it as a hint, not a list.
- 5 Keep looking. New agents, skills and MCP servers appear every week, so a one-off scan is out of date almost at once. Detection has to run all the time, in real time.
How do you reduce shadow AI risk without banning AI?
A ban does not work. Employees who lose one tool move to another, often on a personal account, where security sees even less. The better rule is to say yes to AI and stop only what is dangerous.
Start with what runs. Find every AI tool on every employee device and the person who runs it, then grade each one for risk. Decide on each: allow the low-risk AI, ask before the unclear cases and deny the high-risk ones. That is an allow, ask or deny policy.
Then enforce it where the AI runs, and keep the record. The record is the evidence an auditor asks for. In one AIBound customer result, an unvetted MCP server found on 47 employees' laptops was blocked the same day.
What is an allow, ask or deny policy for AI →Sources
- Microsoft and LinkedIn: 2024 Work Trend Index Annual Report · microsoft.com
- NIST AI RMF Playbook: Govern · airc.nist.gov
Sources checked 19 November 2026
Examples
Shadow AI examples
FAQ
Frequently asked questions
What is shadow AI?
Shadow AI is AI your employees use that security never approved: coding agents, MCP servers, skills, local models and personal-account AI apps. It is common because the tools are easy to install, and it is hard to spot because much of it runs on the device, not on the network.
What are the risks of shadow AI?
Shadow AI can reach company files, code and systems without review. An unvetted skill or MCP server can give an agent more reach than anyone intended, and a personal-account app can take data out of the company. Because nobody knows it is there, it also leaves no evidence for an audit.
What is BYOAI?
BYOAI means bring your own AI: employees use AI tools they chose themselves instead of the ones the company approved. It is another name for the habit behind shadow AI. The approved tools are not the problem. The problem is AI that security cannot see and has not decided on.
How do you detect shadow AI?
Look at the device, because surveys miss what people forget and gateways miss what never reaches the network. AIBound installs on employee devices in 15 minutes and finds the agents, MCP servers, skills and models in real time, with the person who runs each one. You then choose what to allow, ask about or deny.
What are examples of shadow AI?
A coding agent on a developer's laptop that nobody approved. An MCP server that connects an agent to company systems. A skill installed from a public repository. A local model downloaded to summarise reports. An AI app or browser extension signed in with a personal account.
Why do employees use shadow AI?
To get work done faster. The approved tools are often slower, more limited or not there yet, and a new AI tool takes a minute to install. Most employees do not think of it as a security decision, which is why a policy alone does not stop it.
What is the difference between shadow AI and sanctioned AI?
Sanctioned AI is AI that security has reviewed and approved, on a company account, with known access. Shadow AI is everything else. The same tool can be both: an approved coding agent becomes shadow AI when an employee connects it to an MCP server or skill nobody reviewed.
Should you block shadow AI?
Not all of it. Blocking every AI tool pushes employees to personal accounts and devices, where you see less. Find what runs first. Then allow the low-risk AI, ask about the unclear cases and deny the high-risk ones.
How common is shadow AI?
Very common. The 2024 Work Trend Index from Microsoft and LinkedIn found that 78% of AI users bring their own AI tools to work. Because employees install most of it on their own, most companies cannot say how much they have until they look at the device.
How long does shadow AI stay in use before it is found?
Often more than a year. Reco's 2025 State of Shadow AI Report found that unsanctioned AI tools stay in use for more than 400 days on average. Real-time detection on the device shortens that, because a new agent shows up the day it appears.
What tools detect shadow AI?
Three kinds. Network gateways and CASBs read traffic to AI websites. SaaS and expense management tools read invoices and sign-ins. Endpoint tools such as AIBound read the device, where agents, MCP servers, skills and local models run. Much shadow AI never touches the network, so the device view finds what the other two miss.