This document describes how CodexConvert handles user data, API keys, and communication with external services.

CodexConvert is a fully client-side application. There is no backend server, database, or intermediary proxy. All processing happens in the browser:

┌──────────────┐                    ┌──────────────────┐
│  Your Browser │ ══════════════▶  │  AI Provider API  │
│  (CodexConvert)│  direct HTTPS    │  (OpenAI, etc.)  │
└──────────────┘                    └──────────────────┘

CodexConvert itself never receives, stores, or transmits your code or API keys.

Why no backend? A backend server would create a single point where API keys and proprietary code could be intercepted, logged, or stored. By running entirely in the browser, users maintain full control over their data.

File: utils/pathSanitizer.ts

LLM responses contain file paths that are used in ZIP exports. A malicious or hallucinated path like ../../.bashrc could cause a ZIP path traversal attack.

Protection: All LLM-returned paths pass through sanitizePath() which:

1. Normalizes backslashes to forward slashes
2. Splits on /
3. Removes empty segments, ., and ..
4. Rejoins into a safe relative path

This is applied in two layers (defense in depth):

• Required for all remote AI provider URLs
• Exception: HTTP is permitted only for localhost and 127.0.0.1 (for local models like Ollama)
• Remote HTTP URLs are rejected with: "Remote AI providers must use HTTPS."

The validateBaseUrl() function in llmService.ts blocks private network ranges:

Exception: Private/local addresses are allowed when the provider is explicitly set to Ollama.

⚠️ DNS Rebinding Note: DNS rebinding attacks (where a hostname resolves to a public IP during validation but to a private IP at fetch time) cannot be fully prevented in a client-side application. Since this app runs entirely in the user's browser, the impact is limited to the user's own local network.

When a user selects "Custom" as the provider, the UI displays a warning:

⚠️ You are sending your code to a custom AI endpoint. Ensure you trust this provider.

The application includes a CSP meta tag in index.html:

💡 connect-src https: is intentionally broad because CodexConvert must reach whatever AI provider URL the user configures. This is by design.

The entire codebase compiles with "strict": true in tsconfig.json, which enables:

• strictNullChecks — prevents accidental null/undefined access
• noImplicitAny — requires explicit types, reducing runtime surprises
• strictFunctionTypes — enforces correct callback signatures
• strictBindCallApply — prevents incorrect bind/call/apply usage

Before the first conversion in a session, the application displays a privacy notice informing the user:

• Which AI provider and base URL their code will be sent to
• How many models will receive a copy (if multi-model)
• That CodexConvert does not store code or API keys

The user must acknowledge this notice before the conversion proceeds.

If you discover a security vulnerability in CodexConvert, please report it responsibly:

1. Do not open a public GitHub issue for security vulnerabilities
2. Email the maintainer or use GitHub's private vulnerability reporting feature
3. Include:
   • Clear description of the vulnerability
   • Steps to reproduce
   • Potential impact assessment
4. Allow reasonable time for a fix before public disclosure

We aim to acknowledge reports within 48 hours and provide a fix or mitigation plan within 7 days.