Agent identity and permissions

AI agent identity and permissions: who is the agent acting as?

Most AI agents in your company have no identity of their own. They borrow a person's browser session, tokens or terminal and act with everything that person can reach. Before you approve any agent, know whose rights it uses and who can widen them.

Typical agent today

Acting as
The signed-in employee
Can reach
Every site and app with a saved session
Approvals
Set by the user in the agent
Audit trail
Shows the employee, not the agent

Agent you can approve

Acting as
Its own identity with a named owner
Can reach
Only the resources you granted
Approvals
Locked by managed settings
Audit trail
Shows the agent

Next step

AI Agent Blocklist

Block the agents that borrow your staff's identities, such as agent browsers, agent extensions, coding CLIs and agent platforms, and keep only the agents you have given an owner and locked settings.

From $99 a month for one organisation

Four ways an AI agent gets its identity

Every agent needs some identity to act. Ask which of these four it uses, because that decides what it can reach and who sets its limits. Only the last one shows up as itself in your logs.

1. Borrowed browser session

AI browsers and agent extensions

The agent clicks and types inside the user's browser, with the user's cookies and saved logins.

  • Reaches: every signed-in web app
  • Limits set by: the user, in the agent
  • You control: the browser, the extension and the network
2. Delegated grant

Agent platforms and assistants

The user signs in with a work account and consents to access to mail, calendar or files.

  • Reaches: whatever the granted scopes allow
  • Limits set by: consent settings in your identity provider
  • You control: consent rules and the network
3. Developer's shell

Coding agents and CLIs

The agent runs commands as the developer, with their keys, tokens, cloud logins and environment variables.

  • Reaches: repositories, secrets, sometimes production
  • Limits set by: permission and sandbox settings
  • You control: managed settings and installs
4. Registered identity

Agents built on your platform

The agent is registered in your identity system with its own account, an owner and scoped access.

  • Reaches: only what you grant
  • Limits set by: your access policies
  • You control: lifecycle, access and logs
The audit gap: for the first three types, your logs name the employee for every action the agent took. If something goes wrong, you cannot prove whether the person or the agent acted. That gap is why banks and large vendors now refuse agents, as we describe on why companies block AI agents.

AI agent identity security: five permission layers

An agent's own permission prompts are only one layer, and usually the one the user controls. Build from the outside in. Each layer below stops what the layer under it cannot.

1. NetworkCan the agent reach its own service at all? Firewall and DNS rules with agent domains.
2. DeviceMay the app, extension or CLI be installed? Intune, Jamf and browser policies with identifiers.
3. BrowserWhich sites and browser features may an extension touch? Chrome and Edge extension settings.
4. Agent settingsModes, approvals and site lists inside the agent. Lockable only where the vendor offers managed settings.
5. Account rightsWhat the person can do in each app. The agent inherits all of it.
Layers 1 to 3 are yours on every managed device. Layer 4 is yours only when you deploy the vendor's admin settings; otherwise the user decides. Our AI agent governance guide turns these layers into a policy.

Claude in Chrome permissions: a guide for admins

Claude in Chrome is Anthropic's agent extension for Chrome on paid Claude plans, with the ID fcoeoabgfenejglbffodgkkbkcdhcgfn. It acts inside the user's browser, so it is a borrowed session agent with these permission settings.

SettingWhat it doesWho controls it
Manually approveAsks before each action. This mode was called Ask before acting.User, per chat
Automatically approveThe extension checks each action and blocks the ones it judges unsafe.User, per chat
Skip all approvalsNo automatic checks. This mode was called Act without asking.User, per chat
Per-site promptOn a new site: Allow this action, Always allow actions on this site, or Decline.User
Approved sites listUnder Extension settings, Permissions, the user can review and revoke approved sites.User
Enable for your teamTurns the extension on or off for the whole organisation.Admin, Team and Enterprise
Site allowlist and blocklistLimits Claude to approved sites, or blocks sites whatever the user allows.Admin, Team and Enterprise

Always asks, in every mode

  • Changing permission settings
  • Granting authorizations to other services
  • Entering potentially sensitive information into websites
  • Permanent deletions, such as emptying trash or deleting emails and files

Admin controls

  • WhereOrganization settings, then Claude in Chrome, signed in as an Owner or Primary Owner
  • Blocked siteA typed blocked URL shows a banner that the site is blocked by your organisation's policy, and Claude's tools stay off
  • Which accountThe forceLoginOrgUUID policy limits the extension to your organisation
  • PlansAdmin controls exist on Team and Enterprise only, not on individual plans
Why Claude in Chrome keeps asking for permission: the per-site prompt is by design. On company devices we advise against switching to Skip all approvals. Set an admin allowlist of sites instead, or block the extension where you do not want it, as our page on blocking Claude in Chrome shows.

Coding agents: the permission prompts users want to switch off

Searches such as "how to stop Claude Code from asking for permission" come from developers who find the prompts slow. Admins need to know which settings remove the prompts and how to lock them. Both agents below offer managed settings that users cannot override.

Claude Code

  • Modesdefault (shown as Manual), acceptEdits, plan, auto, dontAsk and bypassPermissions
  • What removes promptsThe bypassPermissions mode, which the dangerously-skip-permissions flag also starts. Auto mode replaces prompts with a background check.
  • How to lock itSet permissions.disableBypassPermissionsMode to "disable", and the same for permissions.disableAutoMode if you want every prompt kept
  • Rule orderDeny rules win, then ask, then allow. allowManagedPermissionRulesOnly makes managed rules the only source

Codex CLI

  • Sandbox modesread-only, workspace-write and danger-full-access
  • Approval policieson-request, never and granular; check the current docs, as older values are being retired
  • What removes promptsThe dangerously-bypass-approvals-and-sandbox flag, alias yolo, turns off both the sandbox and approvals
  • How to lock itList the allowed values in allowed_approval_policies and allowed_sandbox_modes in requirements.toml
{ "permissions": { "disableBypassPermissionsMode": "disable", "deny": ["Read(./.env)", "Read(./.env.*)"] } }

Claude Code managed-settings.json. Put it in /Library/Application Support/ClaudeCode/ on macOS, /etc/claude-code/ on Linux and WSL, or C:\Program Files\ClaudeCode\ on Windows.

You can also deliver it as a macOS profile for the com.anthropic.claudecode domain, as a Settings value under HKLM\SOFTWARE\Policies\ClaudeCode, or from the claude.ai admin console.

allowed_approval_policies = ["on-request"] allowed_sandbox_modes = ["read-only", "workspace-write"]

Codex requirements.toml. Put it in /etc/codex/ on macOS and Linux, or %ProgramData%\OpenAI\Codex\ on Windows.

On Macs you can deliver it through MDM in the com.openai.codex domain with the requirements_toml_base64 key, or set it under Agent security in the ChatGPT admin console.

With these files in place, users cannot start the agent with the skip flag, and secrets stay behind deny rules. Read our pages on controlling Claude Code, controlling Codex CLI and keeping agents away from .env files and secrets.

Browser and OS permissions that hand an agent your data

Agent extensions and agent apps ask for broad permissions before they ask for anything inside the agent. These are the Chrome warnings that matter most for agents, with what each one means in practice.

Read and change all your data on all websites

The extension can read and act on every page the user opens, including signed-in apps.

Shown for all-site access and for several powerful APIs
Access the page debugger backend

The extension can drive pages the way automation tools do: click, type and read.

Permission: debugger
Read your browsing history

The extension sees every open tab and its address.

Permission: tabs
Communicate with cooperating native applications

The extension talks to a desktop app, a bridge from the browser to the computer.

Permission: nativeMessaging
Read data you copy and paste

The extension can read copied text, which may include passwords and customer data.

Permission: clipboardRead
Screen Recording and Accessibility

On macOS, agent apps that see and control the screen ask for these in Privacy and Security settings.

Watch for both on the same app

Chrome controls for extension permissions

The Chrome policy ExtensionSettings lets you limit what any extension may do, not only which extensions may install. Test each change on a small group first, because some approved extensions also use these permissions.

ExtensionSettings fieldEffectUse it to
installation_modeblocked stops installs; removed also removes an installed extensionRemove named agent extensions by ID
blocked_permissionsExtensions that request a listed permission cannot install; installed ones stop loadingStop extensions that ask for debugger or native messaging
runtime_blocked_hostsExtensions cannot interact with the listed sites, up to 100 entriesKeep every extension off your bank portal and admin consoles
runtime_allowed_hostsExceptions to the blocked hosts listLet one approved extension work on one blocked site
Agent browsers such as BrowserOS, Comet or ChatGPT Atlas carry the agent inside the browser itself. Policies you set for Google Chrome do not reach a different browser, so block the browser app as well. See blocking Genspark, Fellou and BrowserOS and our Chrome extension allowlist guide.

AI agent identity governance in six steps

Identity management for agents starts with an honest list. These steps work with the identity provider, MDM and firewall you already run.

  1. List every agent

    Record each agent, who runs it and which of the four identity types it uses. Our AI agent inventory guide shows where to look.

  2. Name an owner

    Every approved agent needs a person who answers for it, a stated purpose and a review date. No owner means no approval.

  3. Give it its own identity

    Where the platform supports it, register the agent with its own account instead of a person's. Agents built on Microsoft platforms, for example, can be registered with Microsoft Entra Agent ID.

  4. Require admin consent

    Set your identity provider so users cannot grant new apps access to mail and files on their own. Review the grants that already exist.

  5. Lock the agent's settings

    Deploy managed settings for the agents you allow: site lists for Claude in Chrome, managed settings for Claude Code, requirements for Codex.

  6. Block everything else

    Agents that cannot be identified or locked stay out. Our list covers their domains, extension IDs, app IDs and CLI packages, and it updates every day.

How the AI Agent Blocklist closes the identity gap

You cannot give an identity to an agent you do not know about. The list names the agents that borrow identities, so you can block them on the network and devices while you approve the few that earn an owner.

Where you enforceSetting or toolWhat our file supplies
Firewall and DNS filterExternal or custom block list2,600+ AI agent domains, including agent browsers, agent platforms and MCP servers
Chrome and EdgeExtensionInstallBlocklist or ExtensionSettingsExtension IDs of agent extensions
MacsJamf or Intune restrictionsmacOS bundle IDs and process names of AI browsers and agent apps
WindowsIntune or your endpoint toolsWindows package IDs and process names
Developer machinesPackage proxy or endpoint rulesCommand-line coding agents and their package names
Claude in Chrome: fcoeoabgfenejglbffodgkkbkcdhcgfn Manus AI Browser Operator: cecngibhkljoiafhjfmcgbmikfogdiko Codex CLI, winget: OpenAI.Codex Codex CLI, npm: @openai/codex

A few example entries. Every entry in the full list carries the tool name, its category, the identifier type and a risk grade.

Block

  • Agent browsers on managed devices
  • Agent extensions with no owner
  • Coding agents you have not approved
  • Agent platforms that ask for mail and file access
  • Unapproved MCP servers, see blocking MCP servers

Watch

  • DNS lookups to agent domains
  • New consent grants to agent apps
  • Extensions asking for debugger or native messaging
  • Apps asking for Screen Recording and Accessibility

Allow

  • Agents with their own identity and an owner
  • Coding agents with locked managed settings
  • Claude in Chrome with an admin site allowlist, if approved
  • Read-only assistants that do not act

Keep borrowed-identity agents out

Load the domains into your firewall or DNS filter and the identifiers into your browser and device policies. New agents are added every day, and after an order we set up your download access within one business day.

  • Plain domain list, CSV and JSON with all columns
  • Hosts file, PAC file and firewall external list format
  • DNS filter import files
  • Chrome and Edge extension block lists for Group Policy, Intune and the Google Admin console
  • macOS and Windows identifiers for Jamf and Intune
  • A risk grade on every entry, to block all agents or only the higher-risk ones

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.

Questions about AI agent identity and permissions

What is AI agent identity security?

It is the practice of knowing whose identity an AI agent uses, what that identity can reach, and who can change the agent's permissions. Most agents borrow an employee's browser session, OAuth tokens or terminal, so the audit trail shows the employee.

Agent identity security means giving approved agents their own identity and blocking the agents that cannot be identified.

How do I stop Claude Code from asking for permission?

On company machines, do not remove the prompts with bypass mode. Pre-approve safe, repeated commands with allow rules in managed settings and keep deny rules for secrets.

Set permissions.disableBypassPermissionsMode to disable so the skip flag stops working. Users then see fewer prompts without losing the checks that matter.

How do I stop Codex from asking for permission?

Codex asks for approval when a command needs to leave its sandbox or reach the network. Admins should not hand out the flag that removes both the sandbox and approvals.

Instead, set the allowed approval policies and sandbox modes in requirements.toml so users can work in the workspace without prompts but cannot switch the sandbox off.

Why does Claude in Chrome keep asking for permission?

It asks per site by design. On a new site it shows a prompt with Allow this action, Always allow actions on this site and Decline. Even on approved sites it still asks before downloading files, entering sensitive information and granting authorizations.

On Team and Enterprise plans, admins can limit Claude to an allowlist of sites, which keeps prompts to the sites you approved.

What are the Claude in Chrome permission modes?

The extension offers Manually approve, which was called Ask before acting, Automatically approve, where the extension checks each action and blocks unsafe ones, and Skip all approvals, which was called Act without asking.

Some actions need explicit permission in every mode, such as granting authorizations and permanent deletions.

Is there an AI agent identity management tool from Microsoft?

Microsoft Entra Agent ID gives agents registered in Microsoft Entra their own identities, with owners, access policies and sign-in logs. It covers agents that are registered there.

An extension or command-line agent that an employee installs alone never registers, so you still need device and network rules for those.

Why does an AI agents inventory say a required permission is missing?

Inventory views in admin consoles read data through admin roles and API permissions. When the account you use lacks the role or consent the view needs, it shows a missing permission error or an empty list.

Check which role the console asks for, then sign in with an account that holds it.

Anthropic, Claude, Claude Code and Claude in Chrome are trademarks of Anthropic PBC. OpenAI, Codex, ChatGPT and ChatGPT Atlas are trademarks of OpenAI.

Microsoft, Microsoft Entra, Intune, Edge and Windows are trademarks of Microsoft Corporation. Google and Chrome are trademarks of Google LLC. Apple and macOS are trademarks of Apple Inc. Jamf is a trademark of Jamf.

Perplexity and Comet are trademarks of Perplexity AI, Inc. Manus, BrowserOS, Genspark and Fellou 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.