Permissions Explained

When you install Cogumi AI Shield, Chrome shows a permissions prompt. Here's exactly what we need and why — in plain English.

Why Chrome Shows Permission Warnings

Chrome extensions can request powerful capabilities (reading all websites, accessing your webcam, etc.). Google shows warnings to help you make informed decisions.

Our philosophy: Request only the minimum permissions needed, explain exactly what they're for, and never abuse them.

Permissions We Request

1. Storage

Chrome says: "Read and change your data in chrome.storage"

What we actually do: Store your policy settings, grant history, and audit logs locally on your device (using chrome.storage.local).

What we DON'T do:

Why we need it: Without storage, your security preferences would reset every time you restart Chrome.

Example data stored:

{
  "policy": {
    "global": { "allowPII": false, "allowSecrets": false },
    "chatgpt.com": { "allowPII": true }
  },
  "auditLog": [
    {
      "timestamp": "2026-01-15T10:30:00Z",
      "agent": "ChatGPT",
      "action": "paste",
      "detection": "API key",
      "decision": "denied"
    }
  ]
}

2. Tabs

Chrome says: "Read your browsing history"

Misleading name. We don't read your browsing history. Here's what we actually use this for:

What we do: Check the current tab's URL to determine if you're on an AI agent site (ChatGPT, Claude, etc.).

What we DON'T do:

Code example (from content.ts):

// Only activate on AI agent domains
const AI_AGENTS = [
  'chatgpt.com',
  'claude.ai',
  'bard.google.com',
  'copilot.microsoft.com',
  // ...
];

const currentDomain = new URL(window.location.href).hostname;
if (!AI_AGENTS.includes(currentDomain)) {
  return; // Extension does nothing on non-AI sites
}

Why we need it: To apply the right policy (you might trust your company's internal chatbot but not ChatGPT).

3. Host Permissions (https://*/*)

Chrome says: "Read and change all your data on all websites"

Extremely misleading. This is the scariest permission, but here's the reality:

What we do: Inject a content script that intercepts paste events only on AI agent sites.

What we DON'T do:

Full list of domains where we activate:

Why we request https://*/* instead of specific domains:

Chrome requires "broad host permissions" for extensions that use content scripts. If we listed only specific domains, the extension would break when:

Mitigation: Our code explicitly checks the domain before activating. You can inspect the extension's behavior using Chrome DevTools (Network tab shows zero external requests).

4. Idle

Chrome says: "Read and change your idle state"

What we do: Detect when you're away from your computer to auto-revoke time-limited grants.

Example:

  1. You allow ChatGPT to access API keys for "10 minutes"
  2. You walk away from your desk for 15 minutes
  3. The idle detector sees you've been inactive
  4. The grant is revoked (next paste will re-prompt)

What we DON'T do:

chrome.idle.queryState(60, (state) => {
  if (state === 'locked' || state === 'idle') {
    revokeExpiredGrants(); // Clear old permissions
  }
});

Why we need it: Without this, a "10-minute grant" would last forever if you never closed Chrome.

Permissions We DON'T Request

To demonstrate our commitment to privacy, here are powerful permissions we deliberately avoid:

Cookies

We don't read or modify cookies (no session hijacking).

WebRequest (Network Interception)

We don't intercept network traffic (no HTTPS decryption, no MITM).

Clipboard Read (Unrestricted)

We only read the clipboard when you paste (event-driven, not constant polling).

Downloads

We don't access your downloads (except when you export audit logs, which uses Chrome's standard "Save As" dialog).

Geolocation

We don't care where you are.

Microphone / Camera

Absolutely not.

How to Verify Our Claims

You can monitor the extension's behavior using Chrome DevTools:

1. Check network activity:

2. Inspect storage:

3. Review permissions:

The extension manifest lists all requested permissions:

{
  "permissions": [
    "storage",
    "tabs",
    "idle"
  ],
  "host_permissions": [
    "https://*/*"
  ]
}

Comparison: Cogumi vs. Traditional Enterprise DLP

Permission Cogumi AI Shield (Agentic Security) Typical Enterprise DLP (Traditional) Why Cogumi Needs Less
Network Access No Yes (intercepts HTTPS) Local-first architecture
All Websites Yes (but only activates on AI agents) Yes We filter by domain in code
Cookies No Yes We don't need session tracking
Browsing History No Yes We only check current tab URL
WebRequest No Yes (MITM proxy) No TLS interception needed

Why traditional DLP tools need more permissions:

Why Cogumi needs less:

Future Permission Requests

If we ever need additional permissions (e.g., for a new feature), we'll:

  1. Announce it in the Options Console (in-app notification)
  2. Explain why (detailed blog post)
  3. Make it optional (disable the feature if you don't approve the permission)
  4. Update this page (maintain transparency)

Promise: We will never add analytics, telemetry, or network permissions without explicit user consent.