Ollama Arena is designed for personal, local-first evaluation workflows in trusted environments. It is NOT designed for multi-tenant production deployments or handling highly sensitive data without additional hardening.

• No cloud exposure: Server binds to 127.0.0.1 by default (localhost only)
• No telemetry: Zero external API calls, no tracking, no analytics
• Data sovereignty: All data stays on user's machine unless explicitly exported
• Optional authentication: Set WEB_CHAT_TOKEN environment variable for bearer token protection

• Risk: User prompts not sanitized before rendering
• Impact: Potential XSS if malicious content in model responses
• Mitigation: Markdown renderer escapes HTML, but not foolproof
• Status: Low priority for single-user local deployment

• Risk: No rate limiting on /api/chat or /api/stream_chat endpoints
• Impact: Local DoS possible (flood with requests)
• Mitigation: Single-user deployment reduces risk
• Status: Consider adding for production deployments

• Risk: File upload accepts any file type
• Impact: Users could upload binary files, large files (>10MB limit exists)
• Mitigation: Client-side file reading only, no server-side storage
• Status: Low risk, consider restricting to .txt, .md, .json

• Risk: Conversations stored unencrypted in browser localStorage
• Impact: Anyone with browser access can read all chat history
• Mitigation: localStorage is designed for non-sensitive, personal use
• Status: By design for simplicity; see "Scope" below

• Risk: WEB_CHAT_TOKEN stored in environment variables (plaintext)
• Impact: Token visible to anyone with shell access
• Mitigation: Use OS keychain or secrets manager for production
• Status: Acceptable for local development

• Risk: No CORS restrictions, no CSRF tokens
• Impact: Other localhost apps could make requests
• Mitigation: Requires malicious local app, low risk
• Status: Low priority for trusted local environment

• Personal research: Individual researchers evaluating models on their laptop
• Proof of concept: Testing Ollama integration before production deployment
• Educational: Learning about local LLM orchestration
• Exploratory analysis: Comparing models on non-sensitive datasets

• Multi-user production: Requires user authentication, session management, backend DB
• Regulated data: HIPAA/SOX/GDPR compliance requires encryption at rest, audit logs
• Public deployment: Exposing to internet requires reverse proxy, rate limiting, WAF
• Shared workstations: localStorage is per-browser, no isolation between OS users

If you need to deploy Ollama Arena in a production or shared environment, consider:

• Replace bearer token with proper user authentication (Flask-Login, OAuth)
• Add session management with secure cookies (HttpOnly, SameSite=Strict)
• Implement role-based access control (admin, user, viewer)

• Move from localStorage to backend database (SQLite, PostgreSQL)
• Encrypt sensitive data at rest (SQLCipher, application-level encryption)
• Add audit logging (who accessed what, when)

• Deploy behind reverse proxy (nginx, Caddy) with HTTPS
• Add rate limiting (flask-limiter, nginx limit_req)
• Implement CORS restrictions for API endpoints
• Add CSRF protection for state-changing operations

• Sanitize user prompts before rendering (DOMPurify on client-side)
• Validate file uploads (MIME type, size, content)
• Add Content Security Policy (CSP) headers

• Set up logging aggregation (Elasticsearch, Splunk)
• Monitor for anomalous usage patterns
• Define incident response procedures
• Regular security audits and penetration testing

If you discover a security vulnerability, please:

1. Do NOT open a public GitHub issue
2. Email the maintainer directly: [your-email@example.com]
3. Include:
   • Description of the vulnerability
   • Steps to reproduce
   • Potential impact
   • Suggested fix (if any)

We will respond within 72 hours and coordinate a fix timeline.

• Critical vulnerabilities: Patched within 7 days, immediate release
• High severity: Patched within 30 days
• Medium/Low severity: Included in next minor version

Last updated: January 2026
Security point of contact: [Maintainer Name]