What's changed: Initial: 4 sections for Domain 3 (AI safety, security, and governance)
3.2Data Security and Privacy Controls
Learn data protection for FM deployments: isolation/access with VPC endpoints/IAM/Lake Formation, PII detection with Comprehend/Macie, data masking/anonymization, and retention with S3 Lifecycle.
Because GenAI handles sensitive data, network isolation, least privilege, and PII protection are essential. Ensure data does not leak externally during FM interactions.
3.2.1Protected AI environments
- VPC endpoints: keep traffic to Bedrock private, avoiding the internet.
- IAM/Lake Formation: least-privilege data access and fine-grained (row/column) data governance.
- PII detection: identify personal data with Amazon Comprehend (PII entities) and Amazon Macie (sensitive data discovery in S3).
- Masking/retention: data masking/anonymization and retention/deletion via S3 Lifecycle; combine with Guardrails PII filtering.
Common: private traffic to Bedrock = VPC endpoints, discover sensitive data in S3 = Macie, detect PII in text = Comprehend, row/column data governance = Lake Formation, retention/deletion = S3 Lifecycle. By default, Bedrock does not use customer data to train models.
Privacy protection enforces “do not exfiltrate, do not over-expose, do not over-retain.” Make traffic private with VPC endpoints (PrivateLink) to Bedrock/services, restrict data access with IAM least privilege and Lake Formation row/column controls, and monitor access with CloudWatch. Handle PII on both ingress and egress: detect PII in text with Amazon Comprehend, discover sensitive data in S3 with Amazon Macie, and mask output with Bedrock data privacy features or Guardrails PII filtering. Govern stored data with S3 Lifecycle retention/auto-deletion, and reduce re-identification risk via anonymization/pseudonymization. Note Bedrock does not use prompts/responses to train models by default. In short, combine “private path + least privilege + PII detection/masking + retention policy” to preserve utility while protecting privacy.
| Need | Use | Point |
|---|---|---|
| Private traffic | VPC endpoints | Avoid the internet |
| Discover sensitive data in S3 | Amazon Macie | Auto-discover/classify PII |
| PII in text | Amazon Comprehend | Entity extraction |
| Retention/deletion | S3 Lifecycle | Time-based auto-deletion |
Trap: “data sent to Bedrock is used to train models by default, so you cannot send sensitive data” is wrong—Bedrock does not use customer prompts/responses for training by default (though you still need VPC endpoints, least privilege, PII masking). Also “Macie detects PII in text” is wrong—Macie discovers sensitive data in S3; text PII is Comprehend.
3.2.2Section summary
- Isolation = VPC endpoints / governance = IAM + Lake Formation / PII = Comprehend (text) / Macie (S3)
- Output masking = Guardrails PII / retention = S3 Lifecycle
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Due to regulations, you must keep traffic from the app to Bedrock private, off the internet. Best option?
Q2. You want to automatically discover and classify whether your RAG-source S3 buckets contain sensitive personal data. Best option?
Q3. You want to detect and mask personal data (names, addresses) in prompts/responses to/from the FM. Which fits PII detection in text?