Security· 8 min read

Shadow AI on Company Laptops: How to Find Unapproved AI Tools on Windows

Most small businesses have more AI tools running on their laptops than IT knows about: chat apps, browser extensions, and full AI models running locally. This guide explains what shadow AI is, why it matters, how to find it on Windows with PowerShell, and how to set a simple policy so people can use AI safely.

Rutuja Mali
By · Founder & Managing Director, Sentrivox
Builds SentriOps to give small teams one clear view of their IT assets and security risk. LinkedIn

What shadow AI is

Shadow AI is any AI tool used for work that IT has not approved, or does not know about. On a typical Windows laptop it falls into four groups:

  • AI desktop apps. ChatGPT, Claude and Perplexity desktop apps, installed from the vendor site or the Microsoft Store.
  • Local model runners. Ollama, LM Studio and GPT4All download AI models and run them on the laptop itself, with a small server that other apps talk to.
  • AI browser extensions. Writing assistants, meeting note-takers and page summarisers that can read what is on screen.
  • AI features inside other apps. Assistants built into tools you already approved, sometimes switched on by default.

Why it matters

  • Company data leaves the company. People paste customer lists, contracts, salary sheets and source code into AI chats to get work done faster. Under India's DPDP Act your organisation remains responsible for personal data it processes, wherever an employee sends it.
  • New services appear on the network. Local model runners start a server on the laptop. Ollama listens on port 11434 and LM Studio on port 1234. Ollama binds to the local machine only by default, but one setting change (OLLAMA_HOST=0.0.0.0) makes it reachable by every device on the same network, without a password.
  • Unpatched software. AI tools have security flaws like any other software. Ollama versions before 0.1.34 carried CVE-2024-37032, a path-handling flaw rated High (8.8). Tools that staff install themselves are rarely kept up to date.
  • No audit trail. If an auditor or customer asks which AI tools handle your data, you cannot answer for tools you do not know exist.

Step 1: Find installed AI apps

This lists desktop programs whose names match common AI tools. Run it in PowerShell; it does not need admin rights for the machine-wide entries.

PowerShell
$paths = 'HKLM:\Software\Microsoft\Windows\CurrentVersion\Uninstall\*',
         'HKLM:\Software\WOW6432Node\Microsoft\Windows\CurrentVersion\Uninstall\*',
         'HKCU:\Software\Microsoft\Windows\CurrentVersion\Uninstall\*'
Get-ItemProperty $paths -ErrorAction SilentlyContinue |
  Where-Object DisplayName -match 'Ollama|LM Studio|ChatGPT|Claude|Perplexity|GPT4All' |
  Select-Object DisplayName, DisplayVersion, Publisher

Apps from the Microsoft Store are listed separately. This one needs an elevated window to see every user on the machine:

PowerShell (run as administrator)
Get-AppxPackage -AllUsers |
  Where-Object Name -match 'ChatGPT|Claude|Perplexity' |
  Select-Object Name, Version
Per-user installs
Many AI apps install for one user only. The HKCU line above only covers the account you run it from, so run the check as each user, or use a tool that reads every profile.

Step 2: Check for running AI processes

PowerShell
Get-Process |
  Where-Object ProcessName -match 'ollama|LM Studio|lms|gpt4all' |
  Select-Object ProcessName, Id, Path

A model runner that is installed but never started is lower risk than one running all day as a background service.

Step 3: Check whether an AI service is open to the network

This is the most important check. It shows whether anything is listening on the usual AI ports, and on which address.

PowerShell
Get-NetTCPConnection -State Listen -LocalPort 11434, 1234, 8888 -ErrorAction SilentlyContinue |
  Select-Object LocalAddress, LocalPort, OwningProcess
LocalAddressWhat it meansAction
127.0.0.1 or ::1Only this laptop can reach the serviceLower risk. Keep it updated.
0.0.0.0 or ::Every device on the same network can reach itChange it back to local-only, or block the port in Windows Firewall.
A specific LAN addressReachable on that network interfaceTreat as exposed unless it is meant to be shared.

For Ollama, also check whether someone has changed the listen address:

PowerShell
[Environment]::GetEnvironmentVariable('OLLAMA_HOST', 'User')
[Environment]::GetEnvironmentVariable('OLLAMA_HOST', 'Machine')

Port 8888 is the default for Jupyter notebooks, which are common on data and analytics teams and also run without a login when started carelessly.

Step 4: Check browser extensions

Extensions do not show up in the installed-programs list. Open edge://extensions or chrome://extensions on a sample of machines, or use Microsoft Edge or Google Chrome management policies to list and allow-list extensions across the organisation. Pay attention to extensions that ask to read and change data on all websites.

Step 5: Write a one-page AI usage policy

Finding tools is half the job. People need to know what they can use. Keep the policy short enough to read in two minutes:

  • Approved tools, for example the company Microsoft 365 Copilot licence, with how to get access.
  • Not allowed, for example AI browser extensions that are not on the allow-list, and AI services that train on what you send them.
  • Data that never goes into any AI tool: customer personal data, passwords and keys, source code, and financial or HR records.
  • How to ask for a new tool, and who decides.

Quick checklist

  • Search every Windows device for installed AI apps, including Microsoft Store apps.
  • Check for running AI processes and AI services listening on the network.
  • Change any network-exposed AI service back to local-only, or block the port.
  • Update or remove outdated AI tools, starting with any that have known vulnerabilities.
  • Publish a one-page AI usage policy.
  • Repeat every month. New AI tools appear constantly.

How SentriOps helps

Running these commands on every laptop every month does not scale. SentriOps by Sentrivox gives you the same answers for the whole fleet in one place:

  • Software inventory for every device. The SentriOps Windows agent collects installed software from every endpoint. Search the inventory for Ollama, LM Studio, ChatGPT or Claude and see exactly which machines have them, and which version.
  • Open port visibility. The agent also reports listening ports, so a local AI service exposed to the network stands out without logging in to each machine.
  • Known vulnerabilities, ranked. SentriOps matches installed versions against the National Vulnerability Database and ranks findings by real-world risk with EPSS and the CISA KEV list, so an outdated AI tool with a known flaw lands in your worklist.
  • Evidence for audits. Compliance reports show auditors and customers that you track the software on every device.

Your admin removes or updates the tool, and the SentriOps dashboard shows the result on the next check.

Frequently asked questions

What is shadow AI?

Any AI tool people use for work without IT knowing about it or approving it: AI desktop apps, local model runners such as Ollama or LM Studio, AI browser extensions, and AI features switched on inside other apps.

Is it a problem if staff run AI models locally?

It can be. A local model keeps prompts on the machine, which helps with privacy, but the runner is still software that needs updates, and if it is set to listen on the network, other devices on the same network can use it.

How do I find AI tools installed on Windows?

Search the installed-programs list and Microsoft Store packages for known AI app names, check for running AI processes, and check for AI services listening on ports such as 11434 (Ollama) and 1234 (LM Studio). The PowerShell commands in this guide do all three.

Should we just block all AI tools?

Usually not. Blocking tends to push people to personal phones and accounts, where you have no visibility at all. Approve a small set of tools, say clearly what data must never go into them, and check devices regularly.

Sources and further reading

Related reading