Skip to main content

​
Security for AI Applications

AI applications face unique security challenges that traditional tools weren’t built to handle. Oximy protects against these threats through two foundational concepts: Guardrails and Policies.

​
Core Concepts

  • Guardrails
  • Policies
Real-time security controls that analyze requests and responses as they flow through your AI applications.How They Work:
  • Pattern-based detection for structured threats
  • Semantic analysis to understand context and intent
  • Behavioral analysis to spot unusual patterns
  • Contextual validation against your policies
Response Time: Milliseconds

Learn About Guardrails

Deep dive into how guardrails protect your applications

​
How They Work Together

1

Request Arrives

An AI request arrives at your application (via Gateway or Shield)
2

Policy Loads

The project loads its assigned policy, which specifies:
  • Which guardrails are active
  • Enforcement levels for each guardrail
  • Custom detection settings
3

Guardrails Analyze

Active guardrails analyze the request through multiple layers:
  • Pattern matching for known threats
  • Semantic analysis for context understanding
  • Behavioral analysis for unusual patterns
  • Contextual validation against policies
4

Action Taken

Based on policy configuration:
  • BLOCK: Stop the request, return error, log violation
  • WARN: Sanitize content, log violation, allow to proceed
5

Protected Request

Only clean, policy-compliant content reaches the AI model
6

Response Checked

Model responses are analyzed by guardrails before returning
7

Safe Response

Compliant, sanitized data returns to your application
Visual Flow: 
API Request → Project → Policy → Guardrails → Protected Request → AI Model
                                       ↓
                                  [Analysis]
                                  - Patterns
                                  - Semantics
                                  - Behavior
                                  - Context
                                       ↓
Model Response → Guardrails → Protected Response → Your App

​
Protection by Environment

  • Production
  • Internal Tools
  • Development
  • Compliance
Maximum protection for customer-facing applicationsPolicy Configuration:
  • All guardrails enabled
  • BLOCK enforcement for critical threats
  • High detection sensitivity
  • Real-time alerting
  • Comprehensive audit logging
Use For:
  • Public APIs
  • Customer portals
  • Production services
  • Revenue-generating applications
Example:
{
  "policy": "production-strict",
  "enforcement": {
    "prompt_injection": "BLOCK",
    "data_leakage": "BLOCK",
    "jailbreak": "BLOCK",
    "content_moderation": "BLOCK"
  }
}

​
Why This Architecture?

The same guardrails and policies work across Gateway, Guard, and Shield. Configure once, protect everywhere.
Production needs strict enforcement. Development needs visibility. Same guardrails, different policies.
Pre-built policies for HIPAA, PCI, GDPR, and other regulations. Customize as needed for your requirements.
Start with WARN in development. Monitor and tune. Move to BLOCK in production when you’re confident.
Four detection layers (pattern, semantic, behavioral, contextual) provide defense in depth against evolving threats.

​
Next Steps

1

Understand Guardrails

Learn how each guardrail type protects against specific threatsExplore Guardrails
2

Learn About Policies

See how to organize guardrails into reusable configurationsExplore Policies
3

Integrate Gateway

Put it all into practice with Oximy GatewayGateway Quickstart