If you believe you've found a security issue in CodeRoulette, please report it responsibly.

Email: [email protected]

Or use GitHub Security Advisories for private disclosure:

• Server & web app — coderoulette/coderoulette
• Host agent — coderoulette/host

1. Description of the vulnerability
2. Affected component (server, web app, host agent, shared)
3. Steps to reproduce
4. Impact assessment — what can an attacker do?
5. Suggested fix (if you have one)

CodeRoulette's security relies on several layers:

• Outbound-only connections — the host agent opens a WebSocket to the relay server. No ports are exposed on the host machine.
• Isolated working directory — Claude Code runs in a temp directory or user-specified folder, never the home directory.
• Claude Code permissions — Claude Code's built-in permission system prompts the host before running commands or writing files. The host is always the final gatekeeper.
• Open-source and auditable — the host agent is a thin PTY wrapper with minimal scope.

• JWT authentication — all WebSocket connections require a valid JWT token.
• Input validation — prompts (10K char limit), terminal input (1K char limit), and chat messages (2K char limit) are size-bounded.
• Authorization checks — only the driver can send prompts/input, only the host can approve navigator prompts.

• GitHub OAuth — authentication via NextAuth with GitHub provider.
• Security headers — X-Frame-Options, X-Content-Type-Options, Referrer-Policy, X-XSS-Protection.
• Prompt safety warnings — references to sensitive paths (.ssh, .env, credentials) trigger warnings before sending.

• Claude Code's own security model (report to Anthropic)
• Vulnerabilities requiring physical access to the host machine
• Social engineering attacks
• Denial of service via normal usage patterns

CodeRoulette is a community project. There is no bug bounty program. Please still disclose responsibly so we can fix issues quickly. The best way to help is by sending PRs.