Agentic Design

Patterns
🧀

Layered Defense Pattern(LDP)

Multi-layered security architecture implementing the Swiss Cheese Model for AI safety

Complexity: highSecurity & Privacy Patterns

🎯 30-Second Overview

Pattern: Multiple independent security layers (External → Application → Model) creating defense-in-depth

Why: Single layer failures don't compromise system; each layer catches different threat types

Key Insight: Swiss Cheese Model - holes in different layers don't align, preventing complete breach

⚡ Quick Implementation

1External Layer:API gateway with rate limiting & auth
2Application Layer:Content filtering & PII detection
3Model Layer:Constitutional AI & refusal mechanisms
4Monitor:Log decisions at each layer
5Coordinate:Ensure layers complement, not conflict
Example: request → auth_check → content_filter → model_safety → response_validation

📋 Do's & Don'ts

✅Implement fail-safe defaults at each layer
✅Use defense-in-depth with independent failure modes
✅Log security events for audit and improvement
✅Test layers both independently and integrated
✅Document what each layer protects against
❌Rely on a single layer for critical security
❌Create overlapping rules that conflict
❌Skip monitoring of bypass attempts
❌Ignore performance impact of multiple layers
❌Forget to update all layers when threats evolve

🚦 When to Use

Use When

  • • High-stakes AI applications
  • • Regulatory compliance required
  • • Public-facing systems
  • • Handling sensitive data

Avoid When

  • • Low-risk internal tools
  • • Extreme latency constraints
  • • Simple single-purpose bots
  • • Resource-constrained edge devices

📊 Key Metrics

Block Rate
Threats stopped per layer
False Positives
% legitimate requests blocked
Latency Impact
ms added per layer
Bypass Attempts
Attacks reaching deeper layers
Coverage
% of OWASP Top 10 addressed
MTTR
Time to update defenses

💡 Top Use Cases

Financial AI: External (DDoS) → App (fraud detection) → Model (compliance)
Healthcare Bot: External (HIPAA auth) → App (PHI filtering) → Model (medical safety)
Legal Assistant: External (client auth) → App (privilege checks) → Model (ethical bounds)
HR Platform: External (employee verification) → App (PII protection) → Model (bias prevention)
Educational AI: External (age verification) → App (content filtering) → Model (safety guidelines)

References & Further Reading

Deepen your understanding with these curated resources

Academic Papers & Standards

Designing Multi-layered Runtime Guardrails for Foundation Model Based Agents (ArXiv:2408.02205v3, 2024)

NIST AI Risk Management Framework: Generative AI Profile (NIST AI 600-1, July 2024)

Current State of LLM Risks and AI Guardrails (ArXiv:2406.12934, June 2024)

Swiss Cheese Model in Healthcare Safety Systems (Reason, 2000)

Implementation Guides

OWASP Top 10 for LLM Applications 2025 - Layered Security

AWS Well-Architected Framework - Security Pillar Defense in Depth

Google Cloud AI Platform Security Best Practices

Azure AI Security Baseline - Layered Defenses

Tools & Frameworks

Guardrails AI - Open Source LLM Security Framework

NeMo Guardrails by NVIDIA - Programmable Guardrails

LangChain Security Modules - Multi-layer Implementation

MLflow AI Gateway - Security Layer Management

Industry Resources

NIST AI Safety Institute - Agent Security Guidelines

UK AISI Agent Safety Evaluations

Anthropic Constitutional AI Papers - Model Layer Safety

OpenAI Safety Best Practices - Defense in Depth

Contribute to this collection

Know a great resource? Submit a pull request to add it.

Contribute

Patterns

closed

Loading...

Built by Kortexya