5.3. Reflection Checkpoint: Security, Compliance, and Governance for AI Solutions

Key Takeaways

• Securing an AI system is a layered strategy: IAM controls who can access resources (least privilege), encryption (KMS at rest, TLS in transit) protects the data itself, and PrivateLink isolates traffic from the public internet entirely — no single layer is sufficient alone.
• The Shared Responsibility Model applies to AI exactly as it does elsewhere: AWS secures the AI services themselves (infrastructure, patching, service integrity), while you secure your data, model artifacts, IAM configuration, network setup, and — critically — the responsible, unbiased use of the AI itself.
• Data lineage (tracking which dataset trained which model version) and SageMaker Model Cards (a single governance document per model) are what let you actually answer a regulator's "prove this model wasn't trained on prohibited data" — good documentation isn't optional in a professional AI environment.
• ISO 27001 and SOC 2 are the standards a business is typically judged against; AWS Artifact gives you AWS's own compliance reports to build your case on top of, rather than starting from zero.
• CloudTrail, Config, and Audit Manager solve three distinct governance problems: CloudTrail logs who did what, Config monitors whether resources are still configured correctly, and Audit Manager automates evidence collection for a formal audit.
• Data governance means planning lifecycle (S3 Lifecycle policies), residency (choice of AWS Region), and retention (S3 Object Lock) in advance — not retrofitting them after data already exists in the wrong place.

Connecting Forward

You've now covered all five exam domains: AI/ML fundamentals, generative AI fundamentals, foundation model applications, responsible AI, and security/compliance/governance. Phase 6 shifts entirely to exam readiness — strategy for the exam itself and a path for continued learning afterward.

Self-Check Questions

• A regulator asks a financial company to prove its credit-risk model wasn't trained on a prohibited variable like race. Which two things covered in this phase let the company answer that question quickly and definitively?
• Your company must keep EU customer data within the EU and delete it after five years. Which AWS mechanism enforces each of those two requirements, and why is a single tool not enough to cover both?