Quest 48 - Security Architecture Design

Quest 48: Security Architecture Design

hard 30-45 minutes

🎯 Learning Objectives

✅ Design a threat model for an AI application with database and API access
✅ Plan authentication, authorization, and input validation layers before writing code
✅ Define output guardrails that prevent AI from performing dangerous actions
✅ Create an incident response plan outline for security breaches

📖 Concept: Security by Design — Plan Before You Build

ในโลกของ software security สิ่งที่แย่ที่สุดคือ “ยิงก่อนถามทีหลัง” — สร้าง app ก่อนแล้วค่อยนึกถึง security ทีหลัง Security by Design คือการวางแผน security controls ตั้งแต่ขั้นตอน design ไม่ใช่ตอน deploy

AI applications มีความท้าทายเฉพาะตัว: model ที่คุณ create อาจ execute API calls, เข้าถึง database, หรือส่ง email ได้ — และถ้า attacker ควบคุม model input ได้ พวกเขาก็ควบคุม action เหล่านั้นได้เช่นกัน

Think of it like building a bank: you don’t build the vault first and then wonder where to put the cameras. You design the entire security layout — cameras, locks, guard stations, escape routes — before laying the first brick.

⚙️ How It Works

The Threat Modeling Process

1. Identify assets: What are we protecting?
   → User data, API keys, database, model integrity
         ↓
2. Identify threats: What can go wrong?
   → Prompt injection, data exfiltration, privilege escalation
         ↓
3. Identify controls: How do we prevent each threat?
   → Input validation, output filtering, rate limiting
         ↓
4. Document the architecture
   → Diagrams, policies, response plans

AI-Specific Threat Categories

Threat	Description	Example Attack
Prompt Injection	Attacker overrides system prompt	“Ignore previous instructions”
Data Exfiltration	AI leaks sensitive data in response	Model reveals database contents
Privilege Escalation	AI performs actions beyond its scope	Model executes unauthorized API calls
Model Manipulation	Attacker influences model behavior	Adversarial inputs cause wrong outputs
Supply Chain	Vulnerable dependencies	Malicious npm package

The Security Architecture Layers

┌─────────────────────────────────────────┐
│           OUTPUT GUARDRAILS             │ ← What AI must NOT do
├─────────────────────────────────────────┤
│         AUDIT LOGGING                   │ ← Record everything
├─────────────────────────────────────────┤
│      AUTHORIZATION (RBAC)               │ ← Who can do what
├─────────────────────────────────────────┤
│      AUTHENTICATION (JWT + MFA)         │ ← Who are you
├─────────────────────────────────────────┤
│      INPUT VALIDATION                   │ ← What can come in
├─────────────────────────────────────────┤
│      NETWORK SECURITY                   │ ← WAF, rate limiting
└─────────────────────────────────────────┘

💡 Example: Threat Model for AI Chatbot

## Threat Model: AI Customer Support Chatbot

### Assets
- Customer database (PII, order history)
- API keys (payment processor, email service)
- Model integrity (prevent manipulation)

### Threats & Controls

1. **T1: Prompt Injection → Data Exfiltration**
   - Threat: Attacker tricks AI into revealing customer data
   - Control: Output guardrails + PII detection filter
   - Severity: Critical

2. **T2: API Key Exposure**
   - Threat: AI includes API keys in response
   - Control: Never expose keys to model context
   - Severity: Critical

3. **T3: Unauthorized Actions**
   - Threat: AI executes API calls beyond scope
   - Control: Allowlist of permitted actions + approval workflow
   - Severity: High

4. **T4: SQL Injection via AI**
   - Threat: AI generates SQL with user input
   - Control: Parameterized queries only, no dynamic SQL
   - Severity: High

Key insight: Each threat has a specific control — vague security policies don’t stop specific attacks.

⚠️ Common Mistakes

Mistake 1: Security as an afterthought

“We’ll add security before launch” → Security architecture must be designed before code is written. Retrofitting security is expensive and incomplete.

Mistake 2: Overlooking AI-specific threats

“We have standard web security covered” → AI introduces new attack surfaces: prompt injection, data leakage through responses, model manipulation. Standard web security isn’t enough.

Mistake 3: No output guardrails

“The AI only does what we tell it” → Without output guardrails, a compromised AI can leak data, execute unauthorized actions, or cause harm. Always define what AI must NOT do.

Mistake 4: Missing incident response plan

“We’ll figure it out when something happens” → Incident response plans must exist BEFORE incidents. When a breach happens, you need clear steps, not panic.

📝 Knowledge Check

🏋️ Quest: Security Architecture Design

ออกแบบ security architecture สำหรับ AI chatbot ที่เข้าถึง database และ execute API calls — AI มักจะ proposal architecture ที่ลืม threat modeling หรือ output guardrails!

Download ไฟล์เริ่มต้นของ quest:

npx bluebeltdojo download quest-48-security-arch
cd quest-48-security-arch

เปิด problem.js ใน editor ของคุณพร้อมความช่วยเหลือของ AI
เขียน security-arch.md ที่ครอบคลุม:
- Threat model (อย่างน้อย 4 threats ที่เฉพาะเจาะจง)
- Authentication และ authorization controls
- Input validation และ sanitization layer
- Output guardrails (AI ห้ามทำอะไร)
- Audit logging requirements
- Incident response plan outline
ตรวจสอบ solution ของคุณ:
Terminal window
```
node test.js
```

การตรวจสอบ

node test.js

When all tests pass, you will see the completion message.

ส่งคำตอบ

When tests pass, submit your solution:

npx bluebeltdojo submit

ต้องตั้งค่า access code ก่อน: npx bluebeltdojo setup <code>

คำใบ้

เริ่มจาก threat modeling — ระบุ assets, threats, และ controls ก่อนเขียน architecture
อย่าลืม output guardrails — model ที่เข้าถึง database ได้อาจ leak ข้อมูลถ้าไม่มี guardrails
คิดถึง AI-specific threats: prompt injection, data exfiltration, privilege escalation
ถ้าติดขัด ลองนึกว่าถ้าคุณเป็น attacker จะ exploit AI chatbot ยังไง