Where Kasbah works — honestly

We believe in transparency. Here's exactly what we detect, where we detect it, and what we don't.

Browser support

Browser Protection Detection engine Injection timing
Chrome Full 22-pattern detector + 18-moat egress gate document_start (earliest possible)
Edge Full 22-pattern detector + 18-moat egress gate document_start (earliest possible)
Opera Full 22-pattern detector + 18-moat egress gate document_end
Firefox Full 22-pattern detector + 18-moat egress gate document_start
Safari Full 22-pattern detector + 18-moat egress gate document_start
Brave Strong Same as Chrome (Chromium-based) document_start
Arc Strong Same as Chrome (Chromium-based) document_start

All browsers run the same detector.js engine locally. No data ever leaves your device.

AI platforms we protect

Kasbah monitors data sent to these platforms. If you paste or upload sensitive content, we catch it before it reaches the AI.

ChatGPT (OpenAI)
Claude (Anthropic)
Gemini (Google)
Grok (xAI)
Copilot (Microsoft)
GitHub Copilot
DeepSeek
Perplexity
Poe
Mistral
Meta AI
HuggingFace Chat
Google AI Studio
NotebookLM
Cohere Coral
Manus
Cursor
Windsurf
Codeium
Pi (Inflection)
You.com
Google Labs
Open Assistant
LMSYS Chat Arena

New platforms are added with each release. Chrome extension also protects on all websites, not just AI platforms.

What we detect

22 pattern categories, 100+ language variants, 12 layers of defense. All running locally in your browser.

🆔

Identity documents

Passports, national IDs, driver's licenses, birth certificates — 18 languages

💰

Financial data

Credit cards (Luhn-validated), bank accounts, IBANs, wire transfers, invoices

🏥

Medical records

Patient records, prescriptions, psychiatric notes, insurance claims — 12 languages

📊

Tax documents

SSNs, tax IDs, W-2s, 1040s, 1099s, NIF, SIRET — 12 languages

🔑

API keys & secrets

GitHub PATs, AWS keys, OpenAI keys, bearer tokens, .env secrets, connection strings

🔐

Cryptocurrency

BTC/ETH private keys, wallet addresses (legacy + Bech32), seed phrases, PEM keys

🛡️

Insurance

Auto, health, life, home, and property insurance policies and claims

⚠️

Injection attacks

SQL injection, shell commands, prompt injection — catches code exfiltration

📧

Bulk data

Contact lists (5+ emails), database connection strings with credentials

Languages supported

Full detection (content + filenames)

English, French, Spanish, Portuguese, German, Italian, Dutch, Polish, Arabic, Chinese, Russian, Japanese, Korean, Greek, Turkish

Filename detection

All of the above plus: Moroccan Arabic, CIN/CNIE, IBAN/SWIFT variants, local tax forms (DNI, NIF, SIRET, CIF)

Bypass resistance

Attackers try to evade detection. Here's what we handle:

Handled

L33t speak (p@$$w0rd), Unicode homoglyphs (Cyrillic а vs Latin a), Zalgo text, zero-width characters, fullwidth Unicode, smart quotes, exotic whitespace, base64 encoded content, multi-line splitting

Not yet handled

ROT13 encoding, character-by-character drip (one letter per message), image-embedded text (screenshots of documents), audio-to-text transcriptions

What we don't detect

No security tool is perfect. Here's where our current limits are:

Known limitations

  • Images and screenshots — We scan text and filenames, not pixel content. A screenshot of a passport won't trigger detection.
  • Audio and video — We don't transcribe media files. Spoken secrets in voice messages are not caught.
  • Encrypted or compressed content — If data is zipped or encrypted before pasting, we can't inspect it.
  • Context-dependent secrets — A project codename or internal strategy that only you know is sensitive can't be caught by pattern matching.
  • Very short fragments — Text under 5 characters is skipped for performance. A 4-digit PIN alone won't trigger.
  • Novel encoding schemes — While we handle common evasion (l33t, homoglyphs, base64), truly novel encodings may slip through until patterns are updated.

Our commitment

We'd rather tell you what we can't do than pretend we catch everything. Every version improves detection. If you find something we miss, let us know — it makes everyone safer.

How detection works

1. You type or paste — Content enters a text field on any supported AI platform.

2. Egress gate intercepts — Before the browser sends data, our 18-moat gate catches it (fetch, XHR, WebSocket, form submit, beacon, clipboard).

3. Local classification — detector.js runs 22 regex patterns against the text. All processing happens in your browser. Zero network calls.

4. Risk scoring — Each match adds to a risk score. Multi-pattern matches escalate severity.

5. Three-tier response

Silent (risk < 30) — Safe content passes. No interruption.
Warning (risk 30–70) — Inline toast notification. You decide.
Block (risk > 70) — Modal dialog. Clear explanation of what was detected.

6. You choose — Even on blocks, you can override. We inform, you decide.