Yuxi

Yuxi Security Hardening Guide

Least-privilege permissions

• Restrict agent access by role: Assign each user only the knowledge bases and AI skills they need for their job. Prevent sales teams from accessing HR documents or vice versa.
• Separate admin accounts: Never use administrator credentials for daily work. Create limited accounts that can't modify system settings or user permissions.
• Review tenant boundaries monthly: Confirm that each company/department using Yuxi sees only their own data, documents, and conversation history. Test this with a sample query.
• Audit MCP connectors: Before adding any external skill or tool (MCP connector), verify who created it and what systems it can access—disable or remove suspicious ones.

Network isolation

• Deploy on private servers only: Keep Yuxi behind your company firewall, not accessible from the public internet. Use VPN for remote access.
• Encrypt all connections: Enable HTTPS and require encryption for any communication between Yuxi and your databases, document stores, or external services.
• Limit API key exposure: Store credentials (passwords, API keys, tokens) in a secure vault, never in code or config files. Rotate them every 90 days.
• Restrict outbound connections: Configure firewalls so Yuxi can only reach approved external services—block unexpected data transfers to unknown websites.

Monitoring

• Log all document access: Record who viewed, uploaded, or downloaded knowledge base files. Review logs monthly for unusual activity.
• Alert on permission changes: Set up notifications when roles, access levels, or admin settings are modified—investigate unexpected changes immediately.
• Track agent output: Monitor what information the AI returns to users. Flag cases where sensitive data (SSN, credit card numbers, passwords) appears in responses.
• Monitor failed login attempts: Watch for repeated failed logins or accounts accessing Yuxi at odd times or from unfamiliar locations.

Red flags to watch for

• Responses revealing cross-tenant data: If an agent answers questions about other companies' documents or customers, stop using Yuxi immediately and audit permissions.
• Unexplained credential leaks: If API keys, passwords, or tokens are found in logs or agent responses, assume compromise and rotate all secrets.
• Inactive or abandoned MCP connectors: If add-on tools haven't been updated in 6+ months or show no development activity, replace them with maintained alternatives.
• Slow performance or unplanned downtime: Degraded speed or crashes can indicate a security incident, unauthorized heavy usage, or resource exhaustion—investigate before resuming normal operations.