Security Model

Xilos is designed with security as a foundational principle. All processing happens before queries reach LLM providers, giving organizations complete control over their AI traffic.

Authentication

• All API requests require a Bearer token
• Virtual Keys provide per-user or per-team access with budget and rate limits
• Keys can be deactivated instantly without affecting other users

Guardrails

All queries pass through local guardrails before model routing:

• PII Detection — Identifies and masks sensitive data (SSNs, credit cards, emails, etc.)
• Jailbreak Detection — Identifies prompt injection and manipulation attempts
• Content Moderation — Filters harmful or inappropriate content
• All guardrails run locally — no query content is sent to external services for analysis

Restriction Rules

Organizations create restriction rules to enforce security policies:

• Block — Prevent queries matching specific patterns from processing
• Mask — Redact sensitive information while allowing the query to proceed
• Flag — Allow the query but create an audit trail for review

Data Residency

• Configure data residency to ensure prompts and responses stay in the selected region (US, EU, APAC, or On-Premise)
• When enforced, only providers with matching region endpoints are used

Audit Logging

• Every query is logged with full processing context
• Audit trail captures all governance actions, restriction triggers, and security events
• SIEM export forwards events to external security platforms

Encryption

• Data is encrypted at rest in the database
• All API communication uses HTTPS/TLS
• API keys are hashed in storage