Shadow AI agents

Shadow AI agents: find the agents nobody approved

Somewhere in your company an AI agent is reading mail, filling in forms or running commands with a colleague's login. Nobody reviewed it, nobody owns it, and it does not show up in your app list.

Next step

AI Agent Blocklist: stop agents nobody approved

Domains and device identifiers of AI browsers, agent extensions, desktop agents, coding agents and MCP servers, ready for your firewall, DNS filter, MDM and browser policies.

From $99 a month for one organisation

What shadow AI agents are, and why they are not just shadow AI

Shadow AI used to mean a chatbot someone used without asking. That can still send data out, but it waits for the next prompt. A shadow agent keeps going: it plans steps, opens pages, calls tools and acts until the task is done.

Shadow AI (chat)

An assistant answers questions. A person reads the answer and decides what to do.

  • Risk: text and files pasted into a service you never reviewed
  • Reach: whatever the user copies into it
  • Trace: the user did every action
  • Fix: approved assistant plus a domain block for the rest

Shadow AI agent

An agent gets a goal and works through the steps itself, often with the user's logins and tokens.

  • Risk: actions taken in your systems that nobody checked
  • Reach: every app, tab, folder and token the user can open
  • Trace: logs show the user, not the agent
  • Fix: block by domain and by device identifier, then allow a reviewed few

Short definition you can put in a policy: a shadow AI agent is any AI tool that can take actions on behalf of a member of staff and that the organisation has not approved, registered and assigned an owner to.

Where shadow agents hide in a normal company

Most of them install in minutes without admin rights, or run in the cloud with a token the user granted. That is why the usual software review never sees them.

Agent extensions

A side panel in Chrome or Edge that can read the page and click on it. One install per profile, often from the public store.

Reaches: every logged-in tab

AI browsers

A second browser with an agent built in. Users import their passwords and sessions on first launch.

Reaches: imported logins

Desktop agents

Apps that see the screen and drive the mouse and keyboard, or work on local files in the background.

Reaches: the user's whole desktop

Coding agents

Command-line tools and IDE add-ons that edit code and run shell commands in a project folder.

Reaches: source, secrets, shell

MCP servers

Small connectors that give an assistant tools: mail, files, databases, tickets. Added by editing a config file.

Reaches: the systems they connect

Personal agents

Self-hosted assistants reached through chat apps, running on a work laptop or a home server.

Reaches: mail, calendar, chats

Agent platforms

Hosted builders where staff connect a shared mailbox or CRM to an agent that runs on a schedule.

Reaches: OAuth grants to SaaS

Tenant agents

Agents people build inside your own productivity suite and share with colleagues before anyone reviews them.

Reaches: shared files and sites

AI agent sprawl risks: how one install becomes an incident

Sprawl is the state where agents outnumber the people who can name them. The damage rarely comes from one bad tool. It comes from a chain of small, normal steps that nobody sees end to end.

Borrowed identity

The agent acts as the user. Your audit trail cannot tell a human click from an automated one.

Standing access

Tokens and OAuth grants stay valid long after the task. A forgotten agent keeps its keys.

No owner

When the person who set it up leaves, the agent often keeps running and nobody gets its errors.

Untrusted input

Agents read web pages, mail and documents. Text planted there can steer what they do next.

Data in new places

Screenshots, page content and files go to services that never passed your vendor review.

Silent cost

Agents on paid plans or metered APIs spend money per step. Nobody sees the bill until it arrives.

Why this is urgent now: in December 2025 Gartner published advice titled "Cybersecurity Must Block AI Browsers for Now". Agent extensions and desktop agents carry the same kind of risk without a separate browser to spot.

Signs of shadow agents you can check this week

You do not need a new tool to see the first signals. Most of them are already sitting in consoles you pay for.

SignalWhere you see itWhat it usually means
Extensions asking for "read and change all your data on all websites"Browser extension inventory with permission detailsAn agent or assistant that can act inside any tab
A second Chromium browser on managed laptopsSoftware inventory from your MDMAn AI browser with imported sessions
New OAuth consents to unfamiliar apps with mail or file scopesYour identity provider's enterprise app or consent logsA hosted agent platform connected to a work account
DNS lookups to agent platforms from many devicesDNS filter or resolver logsAgents in daily use, often across a team
Global npm or Python installs of agent packagesDeveloper machine inventory, package manager logsCoding agents and local agent frameworks
JSON config files listing MCP serversEndpoint file inventory on developer machinesAssistants wired to internal systems
Agents shared by staff in your Microsoft 365 tenantMicrosoft 365 admin center, AgentsTenant agents built before any review

Shadow AI agent discovery: a 30-day plan

This plan uses tools most organisations already own. Menu names change as vendors update their consoles, so check the current admin console if a path has moved.

Week 1

Pull what your device tools already know

  1. Intune: Apps > Monitor > Discovered apps, then Export. On company-owned devices this lists all apps; Win32 data refreshes every 24 hours, other app types every seven days.
  2. Microsoft Defender: Endpoints > Vulnerability management > Inventories > Browser extensions. It covers Edge, Chrome and Firefox on Windows and needs the Defender Vulnerability Management add-on or standalone licence.
  3. Google Admin console: Devices > Chrome > Reports > Apps & extensions usage. Chrome browser reporting must be on, and data can take up to 24 hours.
  4. Jamf, Kandji or Mosyle: export the installed application inventory for Macs.
Week 2

Look at the network and the tenant

  1. Export 14 days of DNS queries or firewall URL logs and match them against an agent domain list.
  2. In the Microsoft 365 admin center open Agents > All agents > Registry. Note the cards for Agents without owners and Unmanaged agents, then Export the list.
  3. Review recent OAuth consents in your identity provider for apps with mail, file or calendar scopes.
Week 3

Match, grade and talk to owners

  1. Match every app, extension ID, package and domain against the agent list. Keep the tool name, type and risk grade.
  2. For each match, find a person. Ask what the agent does and what it can reach.
  3. Mark each one block, monitor or allow. Write the reason down.
Week 4

Enforce and keep it current

  1. Push the block list to DNS, the firewall and browser extension policies. Add device identifiers to MDM.
  2. Tell staff which agents are approved and how to ask for a new one.
  3. Repeat the match weekly. New agents appear every day, so a one-time scan goes stale fast.

The full method for building and maintaining the list is on our AI agent inventory page.

Two blind spots: shadow AI desktop agents and agents inside Microsoft 365

Shadow AI desktop agents

Desktop agents run as the signed-in user. Network tools see only encrypted traffic to the vendor, and file access looks like the user opening files.

  • Find them by bundle ID, Windows package ID or process name, not by window title
  • Look for apps that requested screen recording or accessibility rights on Macs
  • Check for local services listening on loopback ports on developer machines
  • Block the install with MDM and the cloud side with DNS

Self-hosted personal agents are a common case; see OpenClaw security risks.

Shadow AI in your Microsoft 365 agents

Staff can build agents and share them before you see them. In the admin center these appear as Shared by creator, next to Microsoft, partner and organisation agents.

  • Agents > Settings > User access: All users, No users, or Specific users or groups
  • Agents > Settings > Sharing: controls sharing of agents built with Agent Builder
  • Agent management rules include Block ownerless agents without usage

The full set of switches is on disable Copilot actions and agents.

Keep in mind: a tenant agent registry lists agents that live in your tenant. An extension, an AI browser or a desktop agent that a user installed is not in it. Those need device and network controls.

How we solve it: one list that finds and blocks shadow agents

The AI Agent Blocklist gives you the two things discovery and enforcement both need: the domains agents talk to and the identifiers they leave on devices. Every entry is checked before it enters the list, and the lists update every day.

Hiding placeWhat the list gives youExample entryWhere it goes
Agent extensionsChrome, Edge and Firefox extension IDscecngibhkljoiafhjfmcgbmikfogdiko (Manus AI Browser Operator)Extension block policy in Intune, Group Policy or Google Admin
Personal agentsExtension IDs, package names, process nameskcdjddhmeafeomebliikmbpblkmkfoig (OpenClaw for Chrome)Browser policy and endpoint inventory rules
Desktop agentsmacOS bundle IDs and Windows package IDsim.manus.desktop, OpenClaw.OpenClawCompanionJamf, Kandji, Mosyle, Intune
Coding agentsCommand-line tools with their package names@openai/codex, openclawPackage proxy rules and endpoint detection
Agent platforms and MCPDomains by category with a risk grademanus.im, smithery.ai, crewai.comDNS filter, firewall, secure web gateway

Decide per entry: block, monitor or allow

Block

  • Agents that act in logged-in browsers
  • Desktop agents with screen control
  • Personal agents on work devices
  • Higher risk grades in every category

Monitor

  • Agent platforms a team is piloting
  • Coding agents pending a security review
  • Lower risk grades while you collect usage

Allow

  • Agents with an owner, a contract and a review
  • Tenant agents published by your organisation
  • Tools on your written approved list

Put it to work in four steps

  1. Discover. Match your inventories and logs against the CSV, which carries tool name, category, identifier type and risk grade for every entry.
  2. Block at the network. Load the domain list into your DNS filter or firewall as a custom or external list. See blocking AI with DNS.
  3. Block on devices. Import the extension block lists into browser policy and the bundle and package IDs into MDM. See blocking AI browser extensions.
  4. Keep the gap closed. Refresh the files on a schedule so new agents land in your rules without anyone retyping IDs.
Plain domain listCSV and JSONHosts filePAC fileFirewall external listDNS filter importsChrome and Edge policy filesmacOS and Windows identifiers

Not sure whether to start from a block list or an allow list? Read allowlist or blocklist for AI agents, then write the rule down in an AI agent acceptable use policy so staff know what is approved.

After discovery: keep shadow agents from coming back

Make the approved route easy

  • Publish a short list of approved agents and what each may touch
  • Give staff one form to request a new agent, with a reply within days
  • Assign every approved agent a named owner and a review date
  • Explain blocks on the block page so people know where to ask

Make the shadow route hard

  • Block agent domains at DNS for every network, including VPN clients
  • Block agent extensions by ID and remove those already installed
  • Stop installs of agent apps and AI browsers through MDM
  • Watch for new OAuth consents with wide scopes

Once the picture is clear, set the rules that decide who may run which agent on our AI agent governance page, and see the wider list of AI agent security risks for your risk register.

The AI Agent Blocklist

  • 2,600+ AI agent domains across browsers, autonomous agents, agent platforms, MCP servers and coding agents
  • 5,300+ device identifiers: extension IDs, macOS bundle IDs, Windows package IDs, process names and CLI package names
  • Tool name, category, identifier type and risk grade on every entry
  • Updated every day; download access set up within one business day of your order

Read what is in each file on the AI Agent Blocklist page.

From $99 a month for one organisation See plans Ask a question

Ask us or order a plan

Tell us what you need to block and where your rules live. We reply within one business day.

  • Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
  • Updated every day as new AI browsers, agents and extensions appear
  • Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you

We use your details only to answer you. No newsletters.

Shadow AI agents: questions we get

What is shadow AI?

Shadow AI is any AI tool that staff use for work without approval from IT or security. Most of it is chat: people paste text into an assistant the company never reviewed. The data leaves, but nothing acts on its own.

What are shadow AI agents?

They are unapproved AI tools that take actions, not only answer questions. They click, type, send, buy, run commands or call other systems, usually with the logins of the person who installed them. AI browsers, agent extensions, desktop agents, coding agents and MCP servers are the common forms.

How do I start shadow AI agent discovery?

Pull the inventories you already have: Intune discovered apps, the Defender browser extensions inventory and the Google Admin apps and extensions usage report. Then export the agent list from the Microsoft 365 admin center.

Match DNS or firewall logs against agent domains and decide block, monitor or allow for each finding.

Can Microsoft 365 agents be shadow AI?

Yes. Agents that staff build and share appear in the Microsoft 365 admin center under Agents as Shared by creator, and the registry shows counts for agents without owners and unmanaged agents. Review them like any other unapproved tool.

What is a shadow AI desktop agent?

It is an installed app that can operate the computer, open apps, read files, type and click, or run shell commands, and that nobody approved. It runs with the user's rights, so it can reach everything the user can.

What are the main AI agent sprawl risks?

Sprawl means many agents with standing access and no owner. The main risks are actions you cannot trace to a person, tokens that outlive the project, data copied to services you never reviewed, and agents that read untrusted content and can be steered by it.

Should we block every AI agent?

Most organisations block unapproved agents by default and allow a short list they have reviewed. Our list carries a risk grade per entry, so you can block everything in the agent categories or start with the higher-risk grades and monitor the rest. See how to block AI agents.

Microsoft, Microsoft 365, Intune, Defender and Edge are trademarks of Microsoft Corporation. Google, Chrome and Google Admin are trademarks of Google LLC. Firefox is a trademark of the Mozilla Foundation. Codex is a trademark of OpenAI. Gartner is a trademark of Gartner, Inc.

Jamf, Kandji, Mosyle, Manus and OpenClaw are trademarks of their respective owners. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.