Browser Extension Guide

The Noxys browser extension is your first line of defense against AI-driven data loss. It monitors 15+ AI platforms in real-time, detects sensitive data instantly, and enforces your organization's policies without requiring proxy installation or system-level certificates.

Supported Browsers & Versions

Browser	Minimum Version	Installation	Notes
Chrome	90+	Chrome Web Store or sideload	Recommended for enterprises
Edge	90+	Microsoft Edge Add-ons or sideload	Full feature parity with Chrome
Brave	1.40+	Chrome Web Store (same URL)	Drops some tracking, otherwise identical
Firefox	88+	Mozilla Add-ons or sideload (.xpi)	Full feature parity
Opera	76+	Chrome Web Store or sideload	Community-supported

Not supported: Safari, Internet Explorer, mobile browsers (support coming in v0.2)

Installation Methods

Method 1: Chrome Web Store (Recommended)

Best for organizations—automatic updates, centralized management.

Visit: https://chrome.google.com/webstore/detail/noxys-ai-firewall/YOUR-EXTENSION-ID
Click Add to Chrome
Review the permissions dialog
Click Add Extension
Icon appears in toolbar (top right)

For Edge, visit Microsoft Edge Add-ons with the same extension ID. For Firefox, visit Mozilla Add-ons.

Method 2: Manual Sideload (Development / Self-Hosted)

Use this if you're self-hosting or testing.

Chrome / Edge

Download extension ZIP: https://your-noxys.instance/downloads/extension.zip
Extract to a folder: /path/to/noxys-extension/
Open chrome://extensions/ (Chrome) or edge://extensions/ (Edge)
Enable Developer Mode (toggle, top right)
Click Load unpacked
Select the extracted folder
Icon appears in toolbar

Firefox

Download extension XPI: https://your-noxys.instance/downloads/extension.xpi
Open about:addons in Firefox
Click the gear icon → Install Add-on From File
Select the .xpi file
Click Add

Note: Firefox requires re-signing for some builds. If it says "This add-on is not available for your version," use the sideload method via about:debugging.

Method 3: Enterprise Deployment

For organizations using Group Policy (GPO) or Mobile Device Management (MDM):

Google Chrome (via GPO)

Create GPO with: Software\Policies\Google\Chrome\ExtensionInstallForcelist
Add value: YOUR-EXTENSION-ID;https://chrome.google.com/webstore/detail/noxys-ai-firewall/YOUR-EXTENSION-ID
Apply to your domain
Users get auto-installed extension on next login

Microsoft Edge (via GPO)

Create GPO with: Software\Policies\Microsoft\Edge\ExtensionInstallForcelist
Same value as Chrome
Apply to your domain

Contact your Noxys account manager for your extension ID.

Initial Configuration

After installation, configure the extension to connect to your Noxys backend.

Setup Wizard (First Run)

On first click of the extension icon:

Popup appears asking for Backend URL and API Token
Backend URL defaults to your Noxys instance (auto-detected from your browser history)
Paste your API token (from Dashboard → Settings → API Keys)
Click Save

Manual Configuration

Click the Noxys icon in your toolbar
Click the gear icon (Settings)
Enter:
- Backend URL: Your Noxys instance (e.g., https://noxys.cloud)
- API Token: From Dashboard → Settings → API Keys
- Mode: Strict (enforce policies) or Permissive (log only)
Click Save

Status Indicator

The extension icon changes color based on connection status:

Color	Status	Meaning
Green	Connected	Successfully authenticated and monitoring
Red	Error	Can't reach backend or invalid token
Gray	Offline	No internet connection (local policies still enforced)
Yellow	Warning	Connection unstable or token expiring soon

Hover over the icon to see connection details.

Supported AI Platforms

The extension monitors 15+ AI platforms. Here's the complete list:

Platform	URL	Aliases	Status
ChatGPT	chat.openai.com, chatgpt.com	chatgpt	✅ Yes
Claude (Anthropic)	claude.ai	claude	✅ Yes
Google Gemini	gemini.google.com	gemini	✅ Yes
DeepSeek	chat.deepseek.com	deepseek	✅ Yes
Perplexity	perplexity.ai	perplexity	✅ Yes
Mistral	chat.mistral.ai	mistral	✅ Yes
Microsoft Copilot	copilot.microsoft.com	copilot	✅ Yes
Poe	poe.com	poe	✅ Yes (beta)
HuggingChat	huggingface.co/chat	huggingchat	✅ Yes
Grok (X AI)	grok.com	grok	✅ Yes
Mammouth AI	mammouth.ai	mammouth	✅ Yes
TypingMind	typingmind.com	typingmind	✅ Yes
OpenRouter	openrouter.ai	openrouter	✅ Yes
Windsurf (Codeium)	windsurf.com	windsurf	✅ Yes
Cursor	cursor.com	cursor	✅ Yes

Coming Soon (v0.2):

Ollama (local)
GitHub Copilot Chat (IDE integration)
VS Code Copilot (IDE integration)

Note: Cursor and Windsurf are primarily IDE tools; browser extension catches web-based use only. VS Code plugin support is on the roadmap.

PII Detection Architecture

The extension uses a three-tier detection strategy, with Tier 1 running locally for speed and privacy.

Tier 1: Extension-Based Regex Detection

Runs client-side in your browser — no data sent to backend unless server-side classification is enabled.

Supported Classifications:

Type	Pattern	Examples
EMAIL	RFC 5322 regex	alice@acme.fr, john.smith@company.co.uk
PHONE	Intl formats (E.164)	+33612345678, +1-555-0123
CREDIT_CARD	Luhn algorithm (PAN)	4532-1234-5678-9010
IBAN	ISO 13616 checksum	FR1420041010050500013M02606
FR_NIR	French social security	1 85 07 75 056 045 72
FR_SIRET	French business ID	73282932000074
FR_SIREN	French company ID	732829320

Performance:

Execution time: <10ms per message
Memory footprint: <5MB
Runs only on 15 AI platforms (doesn't slow down other browsing)

Tier 2: Backend-Based NER (Optional)

Deeper classification using Microsoft Presidio Named Entity Recognition.

Additional Types Detected:

Medical terms (patient names, codes, diagnoses)
Legal references (contract clauses, party names)
API keys and secrets
IP addresses (IPv4, IPv6)
Custom patterns

How to Enable:

Go to extension settings → Advanced
Toggle Enable Server-Side Classification
This sends the prompt text to the backend for Tier 2 analysis (30ms latency added)

Privacy Note: With server-side enabled, the original text is temporarily processed on the backend (Presidio + custom models) then immediately deleted. See Security & Privacy for details.

Tier 3: Async SLM Classification

Context-aware classification using smaller language models (async, no impact on user experience).

Use Cases:

Classify ambiguous content (is "smith" a name or tool name?)
Context-dependent sensitivity (medical vs. non-medical contexts)
Custom classification rules per organization

Risk Score Calculation

risk_score = (number_of_detections) / (content_length) × (average_confidence)

Examples:

1 email detected in 500 chars → risk_score ≈ 0.25
3 emails + 2 phone numbers in 100 chars → risk_score ≈ 0.95 (very high)
No detections → risk_score = 0

Detection Flow

How It Works

1. User types in ChatGPT input box
   ↓
2. Extension content script detects DOM change via observer
   ↓
3. Extracts text from input field
   ↓
4. Runs Tier 1 regex patterns (emails, phone, IBAN, etc.)
   ↓
5. If matches found, calculates risk_score
   ↓
6. Calculates SHA-256 hash of content
   ↓
7. Sends to backend:
   POST /api/v1/interactions {
     platform_id: "chatgpt",
     content_hash: "a665a45920...",
     classifications: [{ type: "EMAIL", ... }],
     risk_score: 0.15,
     content_metadata: { url: "https://chat.openai.com", ... }
   }
   ↓
8. Backend evaluates policies (block/coach/log)
   ↓
9. If action = "block", backend sends response
   ↓
10. Extension prevents submission (blocks button click or shows warning)

No Proxy Required

Unlike other security tools, Noxys extension does NOT require:

MITM proxy installation
Corporate certificate installation
VPN integration
System-level interception

It works entirely in the browser sandbox using the WebExtensions API.

Policy Actions

When the extension detects content matching a policy, three outcomes are possible:

Block

The interaction is prevented. Extension blocks the send button or form submission.

User sees: Red banner "This message contains sensitive data. Blocked by policy: Block PII on ChatGPT"
Backend logs: Policy violation with timestamp, user, platform, risk score
Admin sees: Alert in dashboard
Action: User must edit or cancel

Coach

A warning appears, but the message is NOT blocked. User can still send.

User sees: Yellow banner "This message contains sensitive data. Review before sending."
Backend logs: Coaching event with classifications
Admin sees: Alert in dashboard (non-blocking)
Action: User can click "Send anyway" or edit

Log

Interaction is silently logged (no UI). User doesn't see a warning.

User sees: Nothing
Backend logs: Interaction with classifications and policy decision
Admin sees: In Interactions list and Audit Log (after filtering)
Action: User behavior unaffected

Privacy & Security Model

What's Captured

Platform ID (chatgpt, claude, deepseek, etc.)
SHA-256 hash of your message (NOT the message itself)
Classifications detected (PII types, counts)
URL and metadata (model, timestamp)
Your user ID (from JWT token)

What's NOT Captured

Raw prompt text (except for server-side Tier 2, which is immediately deleted)
Response text from the AI
Your password or session tokens
Other browser tabs or history
Any personal files or desktop data

Data Flow (Secure)

Browser Extension (local)
    ↓
    SHA-256 hash + classifications only
    ↓
TLS 1.3 encrypted connection
    ↓
Noxys Backend (encrypted at rest)
    ↓
PostgreSQL (with AES-256 encryption)

All communication is encrypted in transit (TLS 1.3) and at rest (database encryption).

Offline Mode

If the backend is unreachable (network offline, server down), the extension enters offline mode:

Status icon turns gray
Local policies (if cached) are still enforced
Interactions are queued locally
When connection restored, queued interactions are sent in batch
Queue persists for up to 24 hours in extension storage

To test offline mode:

Open DevTools (F12) → Application → Service Workers
Check Offline checkbox
Try using ChatGPT — warning should still appear if cached policies apply

Advanced Configuration

Change Backend URL

For self-hosted or multi-region deployments:

Click extension icon → Settings → Advanced
Change Backend URL to your custom domain (e.g., https://noxys.mycompany.com)
Click Test Connection

Adjust Detection Sensitivity

Click extension icon → Settings → Advanced
Under Classification Sensitivity:
- Strict (default): Requires high confidence matches (>0.9)
- Balanced: Medium confidence (>0.7)
- Permissive: Any match (>0.5)
Save

Lower sensitivity = more false positives but catches edge cases.

Enable Debug Mode

For troubleshooting:

Click extension icon → Settings → Advanced
Toggle Debug Mode
Open DevTools (F12) → Console
Refresh page
You'll see detailed logs of DOM scanning, hash calculation, and API calls

Troubleshooting

Extension Shows Red Icon (Connection Error)

Problem: Can't connect to backend.

Diagnostic Steps:

Click the extension icon
Look at the status message (hover over the icon)
Go to Settings → Advanced
Verify Backend URL is correct (should be your Noxys instance domain)
Verify API Token is valid (copy from Dashboard → Settings → API Keys)
Check your internet connection (try opening Noxys dashboard in browser)

Solution:

If token expired, generate a new one: Dashboard → Settings → API Keys → Generate Key
Paste into extension settings
Click Test Connection button
If error persists, check network tab (F12 → Network) for exact error

Extension Not Detecting AI Interactions

Problem: Dashboard shows no activity even though you're using ChatGPT.

Diagnostic Steps:

Confirm extension is installed and enabled (check toolbar)
Confirm backend is connected (status icon should be green)
Open ChatGPT in the browser where extension is installed
Check extension permissions: Go to chrome://extensions/ → Noxys → Details → Permissions
- Must include: "Read and change your data on chat.openai.com" and similar for other platforms
Open F12 (DevTools) → Console tab
Refresh ChatGPT
Look for errors related to Noxys or API calls

Solution:

If permission missing, uninstall and reinstall extension
Some enterprise policies block WebExtensions; contact IT
For Brave browser, ensure Shields down for ChatGPT (click shield icon → turn off)
Try a different AI platform (e.g., Claude instead of ChatGPT) to isolate the issue

Policy Not Blocking Interaction

Problem: Policy is enabled but message wasn't blocked.

Diagnostic Steps:

Go to Dashboard → Interactions
Find the recent interaction that should have been blocked
Click to view details: platform_id, risk_score, classifications
Go to Dashboard → Policies
Click the policy that should have triggered
Review conditions and compare against the interaction's fields:
- Does platform_id match?
- Is risk_score above threshold?
- Do classifications match the classification_types condition?
Check priority: Is this policy's priority lower (higher precedence) than others?

Solution:

If conditions don't match, edit the policy to be more permissive (or the interaction data is unexpected)
If priority is wrong, adjust so critical policies run first
Confirm policy is enabled (toggle must be on)

Extension Slowing Down Browser

Problem: Performance degrades after installing extension.

Diagnosis:

Noxys uses <5MB memory (extension + background script)
DOM scanning runs only on AI platforms (15 sites)
Regex matching is optimized (<10ms per message)

Potential causes:

Other extensions conflicting
Too many tabs open
Backend under load (API responses slow)

Solution:

Check browser memory: Ctrl+Shift+Esc (Chrome) → Task Manager → Extensions
Disable Noxys temporarily via chrome://extensions/ (toggle off)
If performance improves, Noxys isn't the culprit
Enable back and check API connectivity (Settings → Test Connection)
If backend is slow, contact support

Extension Crashes or Stops Working

Problem: Status icon disappears or "extension has an error" message.

Solution:

Go to chrome://extensions/
Find Noxys
Toggle off then on (force reload)
If still broken, click Remove
Reinstall from Chrome Web Store
Reconfigure with API token

If crash persists, share the error details:

Open chrome://extensions/ → Noxys → Errors
Copy error message
Email support@noxys.eu with browser version and error

Permissions Denied on Some Platforms

Problem: Extension works on ChatGPT but not Claude.

Cause: Browser permissions for content scripts.

Solution:

Go to chrome://extensions/ → Noxys → Details
Scroll to Permissions → Manage permissions
Add claude.ai to the allowed sites
Refresh Claude in browser
Extension should now work

Next Steps

Admin Console Guide — Manage policies and users
Policy Configuration — Create blocking and coaching policies
PII Detection Types — Understand detection tiers and classifications
Admin Console — Common issues and solutions

Need help?

Email: support@noxys.eu
Documentation: /docs/
Status: Check our status page at status.noxys.cloud

Extension version: 0.1.0 | Last updated: 2026-03-20

Supported Browsers & Versions​

Installation Methods​

Method 1: Chrome Web Store (Recommended)​

Method 2: Manual Sideload (Development / Self-Hosted)​

Chrome / Edge​

Firefox​

Method 3: Enterprise Deployment​

Google Chrome (via GPO)​

Microsoft Edge (via GPO)​

Initial Configuration​

Setup Wizard (First Run)​

Manual Configuration​

Status Indicator​

Supported AI Platforms​

PII Detection Architecture​

Tier 1: Extension-Based Regex Detection​

Tier 2: Backend-Based NER (Optional)​

Tier 3: Async SLM Classification​

Risk Score Calculation​

Detection Flow​

How It Works​

No Proxy Required​

Policy Actions​

Block​

Coach​

Log​

Privacy & Security Model​

What's Captured​

What's NOT Captured​

Data Flow (Secure)​

Offline Mode​

Advanced Configuration​

Change Backend URL​

Adjust Detection Sensitivity​

Enable Debug Mode​

Troubleshooting​

Extension Shows Red Icon (Connection Error)​

Extension Not Detecting AI Interactions​

Policy Not Blocking Interaction​

Extension Slowing Down Browser​

Extension Crashes or Stops Working​

Permissions Denied on Some Platforms​

Next Steps​

Supported Browsers & Versions

Installation Methods

Method 1: Chrome Web Store (Recommended)

Method 2: Manual Sideload (Development / Self-Hosted)

Chrome / Edge

Firefox

Method 3: Enterprise Deployment

Google Chrome (via GPO)

Microsoft Edge (via GPO)

Initial Configuration

Setup Wizard (First Run)

Manual Configuration

Status Indicator

Supported AI Platforms

PII Detection Architecture

Tier 1: Extension-Based Regex Detection

Tier 2: Backend-Based NER (Optional)

Tier 3: Async SLM Classification

Risk Score Calculation

Detection Flow

How It Works

No Proxy Required

Policy Actions

Block

Coach

Log

Privacy & Security Model

What's Captured

What's NOT Captured

Data Flow (Secure)

Offline Mode

Advanced Configuration

Change Backend URL

Adjust Detection Sensitivity

Enable Debug Mode

Troubleshooting

Extension Shows Red Icon (Connection Error)

Extension Not Detecting AI Interactions

Policy Not Blocking Interaction

Extension Slowing Down Browser

Extension Crashes or Stops Working

Permissions Denied on Some Platforms

Next Steps