@xdev-asia/xdev-knowledge-mcp 1.0.41 → 1.0.43

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/05-domain-5-security-compliance/lessons/11-bai-11-ai-security-data-privacy-compliance.md

@@ -0,0 +1,279 @@
+---
+id: 019c9619-lt01-d5-l11
+title: 'Bài 11: AI Security, Data Privacy & Compliance on AWS'
+slug: bai-11-ai-security-data-privacy-compliance
+description: >-
+  IAM cho AI/ML workloads. Encryption at rest & in transit.
+  VPC endpoints cho Bedrock/SageMaker. PII detection.
+  AWS compliance programs. Data governance cho AI.
+duration_minutes: 55
+is_free: true
+video_url: null
+sort_order: 1
+section_title: "Domain 5: Security, Compliance & Governance for AI (14%)"
+course:
+  id: 019c9619-lt01-7001-c001-lt0100000001
+  title: 'Luyện thi AWS Certified AI Practitioner (AIF-C01)'
+  slug: luyen-thi-aws-ai-practitioner
+---
+
+<div style="text-align: center; margin: 2rem 0;">
+<img src="/storage/uploads/2026/04/aws-aif-bai11-ai-security-layers.png" alt="AI Security Layers on AWS" style="max-width: 800px; width: 100%; border-radius: 12px;" />
+<p><em>AI Security Layers: Network, Identity, Data Protection và AI-specific Controls trên AWS</em></p>
+</div>
+
+<h2 id="security-overview"><strong>1. AI Security on AWS — Overview</strong></h2>
+
+<p>AI workloads cần bảo mật ở <strong>nhiều layers</strong>: data, model, infrastructure, và application.</p>
+
+<pre><code class="language-text">AI Security Layers:
+┌─────────────────────────────────────────┐
+│  APPLICATION SECURITY                   │
+│  Guardrails, input validation,          │
+│  prompt injection prevention            │
+├─────────────────────────────────────────┤
+│  MODEL SECURITY                         │
+│  Model access control, versioning,      │
+│  model integrity, adversarial defense   │
+├─────────────────────────────────────────┤
+│  DATA SECURITY                          │
+│  Encryption, PII handling, data access  │
+│  control, audit logging                 │
+├─────────────────────────────────────────┤
+│  INFRASTRUCTURE SECURITY                │
+│  VPC, security groups, endpoints,       │
+│  network isolation, IAM                 │
+└─────────────────────────────────────────┘
+</code></pre>
+
+<h2 id="iam"><strong>2. IAM for AI/ML Workloads</strong></h2>
+
+<h3 id="iam-bedrock"><strong>2.1. IAM for Amazon Bedrock</strong></h3>
+
+<table>
+<thead><tr><th>Permission</th><th>What it controls</th></tr></thead>
+<tbody>
+<tr><td><code>bedrock:InvokeModel</code></td><td>Call a specific FM</td></tr>
+<tr><td><code>bedrock:InvokeModelWithResponseStream</code></td><td>Streaming model invocation</td></tr>
+<tr><td><code>bedrock:CreateKnowledgeBase</code></td><td>Create RAG knowledge bases</td></tr>
+<tr><td><code>bedrock:CreateGuardrail</code></td><td>Create safety guardrails</td></tr>
+<tr><td><code>bedrock:CreateModelCustomizationJob</code></td><td>Start fine-tuning jobs</td></tr>
+</tbody>
+</table>
+
+<h3 id="iam-principles"><strong>2.2. IAM Best Practices for AI</strong></h3>
+
+<ul>
+<li><strong>Least privilege</strong>: Grant only required permissions per role</li>
+<li><strong>Separate roles</strong>: Different roles for data scientists, ML engineers, admins</li>
+<li><strong>Resource-based policies</strong>: Restrict access to specific models</li>
+<li><strong>Service-linked roles</strong>: Let AWS services assume roles as needed</li>
+<li><strong>Condition keys</strong>: Restrict by IP, VPC, time, MFA</li>
+</ul>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "How to restrict which foundation models a team can use in Bedrock?" → IAM policy with <code>bedrock:InvokeModel</code> and <strong>resource ARN</strong> for specific models.</p>
+</blockquote>
+
+<h2 id="encryption"><strong>3. Data Encryption</strong></h2>
+
+<h3 id="encryption-rest"><strong>3.1. Encryption at Rest</strong></h3>
+
+<table>
+<thead><tr><th>Service</th><th>Default Encryption</th><th>Custom Key (KMS)</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon Bedrock</strong></td><td>AWS-managed key</td><td>Customer-managed KMS key</td></tr>
+<tr><td><strong>SageMaker training</strong></td><td>AWS-managed key</td><td>Customer-managed KMS key</td></tr>
+<tr><td><strong>SageMaker notebooks</strong></td><td>AWS-managed key</td><td>Customer-managed KMS key</td></tr>
+<tr><td><strong>S3 (training data)</strong></td><td>SSE-S3</td><td>SSE-KMS, SSE-C</td></tr>
+<tr><td><strong>Knowledge Bases vectors</strong></td><td>Depends on vector DB</td><td>KMS encryption supported</td></tr>
+</tbody>
+</table>
+
+<h3 id="encryption-transit"><strong>3.2. Encryption in Transit</strong></h3>
+
+<ul>
+<li>All AWS API calls use <strong>TLS 1.2+</strong> by default</li>
+<li>Bedrock API calls are encrypted in transit</li>
+<li>SageMaker endpoints use HTTPS</li>
+<li>Inter-node training communication: encrypted</li>
+</ul>
+
+<h3 id="kms"><strong>3.3. AWS KMS for AI</strong></h3>
+
+<pre><code class="language-text">Use AWS KMS when you need:
+✓ Control over encryption keys
+✓ Key rotation policies
+✓ CloudTrail audit of key usage
+✓ Cross-account key sharing
+✓ Compliance requirements (HIPAA, PCI-DSS)
+</code></pre>
+
+<h2 id="network"><strong>4. Network Security</strong></h2>
+
+<h3 id="vpc-endpoints"><strong>4.1. VPC Endpoints for AI Services</strong></h3>
+
+<p><strong>VPC endpoints</strong> cho phép truy cập AWS AI services <strong>privately</strong> — traffic không qua internet.</p>
+
+<pre><code class="language-text">Without VPC Endpoint:
+App in VPC → Internet Gateway → Public Internet → Bedrock API
+
+With VPC Endpoint (PrivateLink):
+App in VPC → VPC Endpoint → AWS Private Network → Bedrock API
+             (no internet!)
+</code></pre>
+
+<table>
+<thead><tr><th>Service</th><th>VPC Endpoint Type</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon Bedrock</strong></td><td>Interface (PrivateLink)</td></tr>
+<tr><td><strong>SageMaker Runtime</strong></td><td>Interface (PrivateLink)</td></tr>
+<tr><td><strong>SageMaker API</strong></td><td>Interface (PrivateLink)</td></tr>
+<tr><td><strong>Amazon S3</strong></td><td>Gateway or Interface</td></tr>
+</tbody>
+</table>
+
+<h3 id="sagemaker-network"><strong>4.2. SageMaker Network Isolation</strong></h3>
+
+<ul>
+<li><strong>VPC mode</strong>: Run training/inference inside your VPC</li>
+<li><strong>Network isolation</strong>: No internet access for containers (EnableNetworkIsolation=true)</li>
+<li><strong>Security groups</strong>: Control inbound/outbound traffic</li>
+<li><strong>Private subnets</strong>: No direct internet access</li>
+</ul>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "How to ensure Bedrock API calls don't traverse the public internet?" → <strong>VPC endpoint</strong> (AWS PrivateLink) for Amazon Bedrock.</p>
+</blockquote>
+
+<h2 id="pii"><strong>5. PII Detection & Data Privacy</strong></h2>
+
+<h3 id="pii-services"><strong>5.1. PII Detection Services</strong></h3>
+
+<table>
+<thead><tr><th>Service</th><th>PII Capability</th><th>Data Type</th></tr></thead>
+<tbody>
+<tr><td><strong>Amazon Comprehend</strong></td><td>Detect and redact PII entities</td><td>Text</td></tr>
+<tr><td><strong>Amazon Macie</strong></td><td>Discover PII in S3 buckets</td><td>Files in S3</td></tr>
+<tr><td><strong>Bedrock Guardrails</strong></td><td>Block/anonymize PII in FM I/O</td><td>FM prompts/responses</td></tr>
+<tr><td><strong>AWS Glue DataBrew</strong></td><td>PII detection in data pipelines</td><td>Structured data</td></tr>
+</tbody>
+</table>
+
+<h3 id="pii-types"><strong>5.2. Common PII Types for Exam</strong></h3>
+
+<table>
+<thead><tr><th>PII Type</th><th>Examples</th></tr></thead>
+<tbody>
+<tr><td><strong>Direct identifiers</strong></td><td>Name, SSN, email, phone, passport number</td></tr>
+<tr><td><strong>Financial</strong></td><td>Credit card number, bank account, tax ID</td></tr>
+<tr><td><strong>Health</strong></td><td>Medical record number, health conditions (PHI)</td></tr>
+<tr><td><strong>Location</strong></td><td>Home address, GPS coordinates</td></tr>
+<tr><td><strong>Digital</strong></td><td>IP address, device ID, login credentials</td></tr>
+</tbody>
+</table>
+
+<h2 id="bedrock-security"><strong>6. Amazon Bedrock Security</strong></h2>
+
+<h3 id="bedrock-data"><strong>6.1. Data Privacy in Bedrock</strong></h3>
+
+<ul>
+<li><strong>Data isolation</strong>: Your data is NOT used to train base FMs</li>
+<li><strong>Data stays in region</strong>: Processed in the AWS region you choose</li>
+<li><strong>Custom models</strong>: Fine-tuned models are private to your account</li>
+<li><strong>No data sharing</strong>: Your prompts/responses are not shared with model providers</li>
+<li><strong>Encryption</strong>: All data encrypted at rest and in transit</li>
+</ul>
+
+<h3 id="bedrock-logging"><strong>6.2. Monitoring & Logging</strong></h3>
+
+<table>
+<thead><tr><th>Service</th><th>What it logs</th></tr></thead>
+<tbody>
+<tr><td><strong>AWS CloudTrail</strong></td><td>API calls (who invoked which model, when)</td></tr>
+<tr><td><strong>Amazon CloudWatch</strong></td><td>Model invocation metrics (latency, errors, tokens)</td></tr>
+<tr><td><strong>Bedrock Model Invocation Logging</strong></td><td>Full prompts and responses (to S3 or CloudWatch)</td></tr>
+</tbody>
+</table>
+
+<blockquote>
+<p><strong>Exam tip:</strong> "How to audit which users are calling Bedrock models?" → <strong>CloudTrail</strong>. "How to log the actual prompts and responses?" → <strong>Bedrock Model Invocation Logging</strong>.</p>
+</blockquote>
+
+<h2 id="compliance"><strong>7. Compliance & Governance</strong></h2>
+
+<h3 id="compliance-programs"><strong>7.1. AWS Compliance Programs</strong></h3>
+
+<table>
+<thead><tr><th>Program</th><th>What</th><th>Relevant AI Services</th></tr></thead>
+<tbody>
+<tr><td><strong>SOC 1/2/3</strong></td><td>Security controls audit</td><td>Bedrock, SageMaker</td></tr>
+<tr><td><strong>HIPAA</strong></td><td>Healthcare data protection</td><td>SageMaker, Comprehend Medical</td></tr>
+<tr><td><strong>GDPR</strong></td><td>EU data privacy</td><td>All AWS services (data residency)</td></tr>
+<tr><td><strong>PCI-DSS</strong></td><td>Payment card data security</td><td>SageMaker (with controls)</td></tr>
+<tr><td><strong>FedRAMP</strong></td><td>US government cloud security</td><td>GovCloud regions</td></tr>
+<tr><td><strong>ISO 27001</strong></td><td>Information security management</td><td>Bedrock, SageMaker</td></tr>
+</tbody>
+</table>
+
+<h3 id="shared-responsibility"><strong>7.2. Shared Responsibility Model for AI</strong></h3>
+
+<pre><code class="language-text">CUSTOMER responsibility ("Security IN the cloud"):
+├── Training data quality and bias
+├── Model selection and evaluation
+├── Prompt design and guardrails configuration
+├── IAM permissions and access control
+├── PII handling and data classification
+├── Application-level security
+└── Compliance with industry regulations
+
+AWS responsibility ("Security OF the cloud"):
+├── Physical infrastructure security
+├── Network and hardware security
+├── Base FM provider management
+├── Service availability and reliability
+├── Encryption implementation
+└── Compliance certifications
+</code></pre>
+
+<h2 id="data-governance"><strong>8. Data Governance for AI</strong></h2>
+
+<table>
+<thead><tr><th>Practice</th><th>AWS Service</th></tr></thead>
+<tbody>
+<tr><td>Data cataloging</td><td>AWS Glue Data Catalog</td></tr>
+<tr><td>Data classification</td><td>Amazon Macie</td></tr>
+<tr><td>Access control</td><td>AWS Lake Formation</td></tr>
+<tr><td>Data lineage</td><td>SageMaker ML Lineage Tracking</td></tr>
+<tr><td>Data quality</td><td>SageMaker Data Wrangler, Glue DataBrew</td></tr>
+</tbody>
+</table>
+
+<h2 id="practice-questions"><strong>9. Practice Questions</strong></h2>
+
+<p><strong>Q1:</strong> A financial services company wants to use Amazon Bedrock but requires that API calls do NOT traverse the public internet. What should they configure?</p>
+<ul>
+<li>A) AWS Direct Connect</li>
+<li>B) VPC endpoint (AWS PrivateLink) for Bedrock ✓</li>
+<li>C) VPN connection</li>
+<li>D) CloudFront distribution</li>
+</ul>
+<p><em>Explanation: A VPC interface endpoint (PrivateLink) for Amazon Bedrock allows private connectivity from within a VPC without traffic going through the internet.</em></p>
+
+<p><strong>Q2:</strong> According to the AWS Shared Responsibility Model, who is responsible for ensuring training data does not contain bias?</p>
+<ul>
+<li>A) AWS</li>
+<li>B) The foundation model provider</li>
+<li>C) The customer ✓</li>
+<li>D) Both AWS and the customer equally</li>
+</ul>
+<p><em>Explanation: Under the shared responsibility model, customers are responsible for "security IN the cloud" — this includes training data quality, bias detection, and ethical AI practices. AWS is responsible for infrastructure security.</em></p>
+
+<p><strong>Q3:</strong> A company needs to discover which S3 buckets contain personally identifiable information (PII) before using the data for ML training. Which service should they use?</p>
+<ul>
+<li>A) Amazon Comprehend</li>
+<li>B) Amazon Macie ✓</li>
+<li>C) Amazon Inspector</li>
+<li>D) AWS Config</li>
+</ul>
+<p><em>Explanation: Amazon Macie uses machine learning to automatically discover and classify sensitive data (including PII) stored in Amazon S3 buckets. Comprehend detects PII in text, but Macie is designed for S3-level discovery.</em></p>

package/content/series/luyen-thi/luyen-thi-aws-ai-practitioner/chapters/05-domain-5-security-compliance/lessons/12-bai-12-exam-strategy-cheat-sheet.md

@@ -0,0 +1,229 @@
+---
+id: 019c9619-lt01-d5-l12
+title: 'Bài 12: Exam Strategy & Complete Cheat Sheet'
+slug: bai-12-exam-strategy-cheat-sheet
+description: >-
+  AWS AI Practitioner exam strategy. Time management.
+  Complete cheat sheet cho tất cả 5 domains.
+  Key services mapping. Common exam traps.
+duration_minutes: 45
+is_free: true
+video_url: null
+sort_order: 2
+section_title: "Domain 5: Security, Compliance & Governance for AI (14%)"
+course:
+  id: 019c9619-lt01-7001-c001-lt0100000001
+  title: 'Luyện thi AWS Certified AI Practitioner (AIF-C01)'
+  slug: luyen-thi-aws-ai-practitioner
+---
+
+<h2 id="exam-overview"><strong>1. Exam Overview — AIF-C01</strong></h2>
+
+<table>
+<thead><tr><th>Detail</th><th>Value</th></tr></thead>
+<tbody>
+<tr><td><strong>Exam code</strong></td><td>AIF-C01</td></tr>
+<tr><td><strong>Questions</strong></td><td>65 scored + 15 unscored = 80 total</td></tr>
+<tr><td><strong>Duration</strong></td><td>90 minutes</td></tr>
+<tr><td><strong>Passing score</strong></td><td>700 / 1000</td></tr>
+<tr><td><strong>Cost</strong></td><td>$100 USD</td></tr>
+<tr><td><strong>Format</strong></td><td>Multiple choice + Multiple response</td></tr>
+<tr><td><strong>Delivery</strong></td><td>Pearson VUE (testing center or online proctored)</td></tr>
+<tr><td><strong>Validity</strong></td><td>3 years</td></tr>
+</tbody>
+</table>
+
+<h2 id="domains"><strong>2. Domain Weights</strong></h2>
+
+<pre><code class="language-text">Domain 1: Fundamentals of AI and ML ████████████ 20%  (~13 questions)
+Domain 2: Fundamentals of Generative AI █████████████████ 24%  (~16 questions)
+Domain 3: Applications of Foundation Models ████████████████████ 28%  (~18 questions)
+Domain 4: Guidelines for Responsible AI █████████ 14%  (~9 questions)
+Domain 5: Security, Compliance &amp; Governance █████████ 14%  (~9 questions)
+</code></pre>
+
+<blockquote>
+<p><strong>Strategy:</strong> Domain 3 (28%) là phần lớn nhất — tập trung nhiều nhất vào Bedrock, RAG, Prompt Engineering, Fine-tuning. Domain 1+2 (44%) là foundation — hiểu concepts là đủ.</p>
+</blockquote>
+
+<h2 id="time-management"><strong>3. Time Management Strategy</strong></h2>
+
+<pre><code class="language-text">80 questions / 90 minutes = ~67 seconds per question
+
+Strategy:
+┌─────────────────────────────────────────────┐
+│ Pass 1 (0-60 min): Answer easy questions    │
+│   → Skip difficult ones (flag for review)   │
+│   → Target: 60+ questions answered          │
+├─────────────────────────────────────────────┤
+│ Pass 2 (60-85 min): Review flagged questions│
+│   → Eliminate wrong answers first           │
+│   → Use process of elimination              │
+├─────────────────────────────────────────────┤
+│ Pass 3 (85-90 min): Final review            │
+│   → Never leave questions blank             │
+│   → No penalty for guessing                 │
+└─────────────────────────────────────────────┘
+</code></pre>
+
+<h2 id="cheatsheet-d1"><strong>4. Domain 1 Cheat Sheet — AI/ML Fundamentals</strong></h2>
+
+<table>
+<thead><tr><th>Concept</th><th>Key Points</th></tr></thead>
+<tbody>
+<tr><td><strong>AI vs ML vs DL</strong></td><td>AI ⊃ ML ⊃ DL. DL uses neural networks.</td></tr>
+<tr><td><strong>Supervised</strong></td><td>Labeled data → Classification (discrete) or Regression (continuous)</td></tr>
+<tr><td><strong>Unsupervised</strong></td><td>Unlabeled data → Clustering, dimensionality reduction</td></tr>
+<tr><td><strong>Reinforcement</strong></td><td>Agent + Environment + Rewards → Learn by trial and error</td></tr>
+<tr><td><strong>Overfitting</strong></td><td>Good on training, bad on test → More data, regularization, dropout</td></tr>
+<tr><td><strong>Underfitting</strong></td><td>Bad on both → More complex model, more features, longer training</td></tr>
+<tr><td><strong>Precision</strong></td><td>Of predicted positives, how many are correct? (avoid false positives)</td></tr>
+<tr><td><strong>Recall</strong></td><td>Of actual positives, how many did we find? (avoid false negatives)</td></tr>
+<tr><td><strong>F1 Score</strong></td><td>Harmonic mean of Precision and Recall</td></tr>
+<tr><td><strong>AUC-ROC</strong></td><td>Model's ability to distinguish classes (higher = better)</td></tr>
+</tbody>
+</table>
+
+<h2 id="cheatsheet-d2"><strong>5. Domain 2 Cheat Sheet — Generative AI</strong></h2>
+
+<table>
+<thead><tr><th>Concept</th><th>Key Points</th></tr></thead>
+<tbody>
+<tr><td><strong>Foundation Models</strong></td><td>Large models trained on broad data, adaptable to many tasks</td></tr>
+<tr><td><strong>Transformer</strong></td><td>Encoder-only (BERT, classify), Decoder-only (GPT, generate), Both (T5, translate)</td></tr>
+<tr><td><strong>Tokens</strong></td><td>~4 chars = 1 token. Context window = max tokens. Pricing = per token.</td></tr>
+<tr><td><strong>Temperature</strong></td><td>Low (0) = deterministic/factual. High (1) = creative/diverse.</td></tr>
+<tr><td><strong>Top-p</strong></td><td>Nucleus sampling. 0.1 = focused. 0.9 = diverse.</td></tr>
+<tr><td><strong>Hallucination</strong></td><td>Model generates false info. Mitigate with: RAG, guardrails, lower temp.</td></tr>
+<tr><td><strong>Embeddings</strong></td><td>Text → vectors. Similar meaning = close vectors. Used for search, RAG.</td></tr>
+<tr><td><strong>Diffusion models</strong></td><td>Image gen by denoising. Stable Diffusion, Titan Image Generator.</td></tr>
+<tr><td><strong>Multi-modal</strong></td><td>Process multiple data types. Claude 3 (text+image), Titan MM Embeddings.</td></tr>
+</tbody>
+</table>
+
+<h2 id="cheatsheet-d3"><strong>6. Domain 3 Cheat Sheet — Applications</strong></h2>
+
+<table>
+<thead><tr><th>Concept</th><th>Key Points</th></tr></thead>
+<tbody>
+<tr><td><strong>Zero-shot</strong></td><td>No examples. Model uses its knowledge.</td></tr>
+<tr><td><strong>Few-shot</strong></td><td>2-5 examples. Pattern following.</td></tr>
+<tr><td><strong>Chain-of-Thought</strong></td><td>"Think step by step" → better reasoning/math.</td></tr>
+<tr><td><strong>RAG</strong></td><td>Retrieve docs → augment prompt → generate. Reduces hallucination.</td></tr>
+<tr><td><strong>Bedrock Knowledge Bases</strong></td><td>Managed RAG: S3 docs → auto chunk → embed → vector store → retrieve.</td></tr>
+<tr><td><strong>Fine-tuning</strong></td><td>Further train FM with your data. Need 1000+ labeled examples. LoRA = cheaper.</td></tr>
+<tr><td><strong>Bedrock Agents</strong></td><td>Multi-step tasks: plan → call APIs → return results. Uses action groups.</td></tr>
+<tr><td><strong>Bedrock Guardrails</strong></td><td>Content filter, denied topics, PII filter, word filter, grounding check.</td></tr>
+<tr><td><strong>PartyRock</strong></td><td>Free, no-code Bedrock playground. No AWS account needed.</td></tr>
+<tr><td><strong>Amazon Q Developer</strong></td><td>AI code assistant (was CodeWhisperer). IDE integration.</td></tr>
+<tr><td><strong>Amazon Q Business</strong></td><td>Enterprise Q&A on company data. Respects ACLs.</td></tr>
+</tbody>
+</table>
+
+<h2 id="cheatsheet-d4"><strong>7. Domain 4 Cheat Sheet — Responsible AI</strong></h2>
+
+<table>
+<thead><tr><th>Concept</th><th>Key Points</th></tr></thead>
+<tbody>
+<tr><td><strong>Bias types</strong></td><td>Selection, sampling, measurement, label, algorithmic, confirmation.</td></tr>
+<tr><td><strong>SageMaker Clarify</strong></td><td>Pre-training bias + post-training bias + SHAP explainability.</td></tr>
+<tr><td><strong>Amazon A2I</strong></td><td>Human-in-the-loop. Trigger human review when confidence low.</td></tr>
+<tr><td><strong>SHAP</strong></td><td>Feature importance per prediction. "Why was this denied?"</td></tr>
+<tr><td><strong>AI Service Cards</strong></td><td>AWS public docs: intended use, limitations, fairness for AWS AI services.</td></tr>
+<tr><td><strong>Model Cards</strong></td><td>SageMaker: document YOUR model's details, metrics, ethics.</td></tr>
+<tr><td><strong>Toxicity detection</strong></td><td>Guardrails content filters + Amazon Comprehend toxicity.</td></tr>
+</tbody>
+</table>
+
+<h2 id="cheatsheet-d5"><strong>8. Domain 5 Cheat Sheet — Security & Compliance</strong></h2>
+
+<table>
+<thead><tr><th>Concept</th><th>Key Points</th></tr></thead>
+<tbody>
+<tr><td><strong>IAM</strong></td><td>Least privilege. bedrock:InvokeModel + resource ARN = restrict models.</td></tr>
+<tr><td><strong>Encryption at rest</strong></td><td>KMS keys for Bedrock, SageMaker, S3. Customer-managed for compliance.</td></tr>
+<tr><td><strong>VPC endpoints</strong></td><td>PrivateLink = no internet. Bedrock + SageMaker support this.</td></tr>
+<tr><td><strong>PII detection</strong></td><td>Comprehend (text), Macie (S3), Guardrails (FM I/O).</td></tr>
+<tr><td><strong>CloudTrail</strong></td><td>Who called what API when. Audit trail.</td></tr>
+<tr><td><strong>Model Invocation Logging</strong></td><td>Log actual prompts & responses (to S3/CloudWatch).</td></tr>
+<tr><td><strong>Bedrock data privacy</strong></td><td>Your data NOT used to train base FMs. Stays in region. Encrypted.</td></tr>
+<tr><td><strong>Shared Responsibility</strong></td><td>Customer: data, bias, IAM, compliance. AWS: infra, service security.</td></tr>
+</tbody>
+</table>
+
+<h2 id="service-mapping"><strong>9. "Which AWS Service?" Quick Reference</strong></h2>
+
+<table>
+<thead><tr><th>If you need to...</th><th>Use this service</th></tr></thead>
+<tbody>
+<tr><td>Access multiple FMs via single API</td><td><strong>Amazon Bedrock</strong></td></tr>
+<tr><td>Build RAG with company docs</td><td><strong>Bedrock Knowledge Bases</strong></td></tr>
+<tr><td>Create AI that calls APIs</td><td><strong>Bedrock Agents</strong></td></tr>
+<tr><td>Block harmful/PII content</td><td><strong>Bedrock Guardrails</strong></td></tr>
+<tr><td>Fine-tune FM</td><td><strong>Bedrock Custom Models</strong> or <strong>SageMaker</strong></td></tr>
+<tr><td>Train custom ML model</td><td><strong>Amazon SageMaker</strong></td></tr>
+<tr><td>Auto-ML (no code)</td><td><strong>SageMaker Canvas</strong> or <strong>SageMaker Autopilot</strong></td></tr>
+<tr><td>Detect bias in model</td><td><strong>SageMaker Clarify</strong></td></tr>
+<tr><td>Human review for AI predictions</td><td><strong>Amazon A2I</strong></td></tr>
+<tr><td>Extract text from documents</td><td><strong>Amazon Textract</strong></td></tr>
+<tr><td>Analyze images/video</td><td><strong>Amazon Rekognition</strong></td></tr>
+<tr><td>Translate text</td><td><strong>Amazon Translate</strong></td></tr>
+<tr><td>Speech-to-text</td><td><strong>Amazon Transcribe</strong></td></tr>
+<tr><td>Text-to-speech</td><td><strong>Amazon Polly</strong></td></tr>
+<tr><td>NLP (sentiment, entities)</td><td><strong>Amazon Comprehend</strong></td></tr>
+<tr><td>Build chatbot</td><td><strong>Amazon Lex</strong></td></tr>
+<tr><td>Enterprise search</td><td><strong>Amazon Kendra</strong></td></tr>
+<tr><td>Recommendations</td><td><strong>Amazon Personalize</strong></td></tr>
+<tr><td>Time-series forecasting</td><td><strong>Amazon Forecast</strong></td></tr>
+<tr><td>Fraud detection</td><td><strong>Amazon Fraud Detector</strong></td></tr>
+<tr><td>Code assistant</td><td><strong>Amazon Q Developer</strong></td></tr>
+<tr><td>Enterprise Q&A on company data</td><td><strong>Amazon Q Business</strong></td></tr>
+<tr><td>No-code GenAI playground</td><td><strong>PartyRock</strong></td></tr>
+<tr><td>Find PII in S3</td><td><strong>Amazon Macie</strong></td></tr>
+</tbody>
+</table>
+
+<h2 id="common-traps"><strong>10. Common Exam Traps</strong></h2>
+
+<h3 id="trap-list"><strong>Watch out for these:</strong></h3>
+
+<ol>
+<li><strong>"MOST suitable" / "BEST"</strong>: Multiple answers may work — choose the one that best fits ALL constraints.</li>
+<li><strong>RAG vs Fine-tuning</strong>: If question mentions "latest data" or "company docs" → RAG. If "specific style" or "writing pattern" → Fine-tuning.</li>
+<li><strong>System prompt vs Guardrails</strong>: If question says "guarantee" or "enforce" → Guardrails (system prompts can be bypassed).</li>
+<li><strong>Supervised vs Unsupervised</strong>: "Labels" = supervised. "No labels" = unsupervised. "Reward" = reinforcement.</li>
+<li><strong>SageMaker vs Bedrock</strong>: Custom ML model from scratch → SageMaker. Use/customize existing FM → Bedrock.</li>
+<li><strong>Comprehend vs Macie</strong>: PII in text (runtime) → Comprehend. PII in S3 files (discovery) → Macie.</li>
+<li><strong>Clarify vs Guardrails</strong>: Bias detection → Clarify. Content safety → Guardrails.</li>
+<li><strong>Temperature</strong>: "Consistent answers" → Low temperature. "Creative responses" → High temperature.</li>
+<li><strong>"Minimum effort" / "least overhead"</strong>: Usually means managed service (Bedrock > SageMaker).</li>
+<li><strong>CloudTrail vs CloudWatch vs Invocation Logging</strong>: API audit → CloudTrail. Metrics → CloudWatch. Prompt/response logs → Invocation Logging.</li>
+</ol>
+
+<h2 id="exam-day"><strong>11. Exam Day Tips</strong></h2>
+
+<ul>
+<li><strong>Read the ENTIRE question</strong> before looking at answers</li>
+<li><strong>Identify keywords</strong>: "MOST", "LEAST", "BEST", "FIRST", "guarantee"</li>
+<li><strong>Eliminate wrong answers</strong> first — usually 2 are obviously wrong</li>
+<li><strong>Don't overthink</strong> — if you know the answer, select it and move on</li>
+<li><strong>Flag and skip</strong> hard questions — come back in Pass 2</li>
+<li><strong>Never leave blank</strong> — no penalty for guessing</li>
+<li><strong>15 unscored questions</strong> — you don't know which ones, so answer all seriously</li>
+<li><strong>Manage time</strong> — 67 seconds per question average</li>
+</ul>
+
+<h2 id="next-steps"><strong>12. Next Steps</strong></h2>
+
+<ol>
+<li><strong>Review all 11 lessons in this series</strong></li>
+<li><strong>Take the mock exam</strong> in the "Thi thử" section</li>
+<li><strong>AWS Skill Builder</strong>: Free "Exam Prep" course for AIF-C01</li>
+<li><strong>AWS PartnerCast</strong>: Additional exam prep resources</li>
+<li><strong>Practice with Bedrock</strong>: Use AWS Free Tier / PartyRock</li>
+<li><strong>Schedule the exam</strong> at <a href="https://aws.amazon.com/certification/">aws.amazon.com/certification</a></li>
+</ol>
+
+<blockquote>
+<p><strong>Remember:</strong> Bạn không cần hands-on experience — đây là Practitioner level exam. Tập trung vào <strong>concepts, use cases, và chọn đúng AWS service</strong>.</p>
+</blockquote>