AI Safety, Security, and Governance
Data Security, Privacy, and Least Privilege
Protect prompts, responses, embeddings, logs, and source data with least privilege, private networking, encryption, data classification, masking, retention controls, and tenant isolation.
Concepts
- Apply IAM policies at the model, knowledge source, tool, and data-resource levels instead of granting broad service access.
- Use KMS keys and private connectivity where compliance or threat models require customer control and network isolation.
- Detect and redact sensitive data before inference and before logs or outputs cross trust boundaries.
- Enforce tenant filters in retrieval and storage layers; prompt instructions are not an access-control mechanism.
- AgentCore Policy can enforce Cedar authorization over Gateway tool calls using trusted principals and arguments, but target services still validate authoritative business state.
Exam tips
- Amazon Macie discovers sensitive data in S3; Amazon Comprehend can detect PII in text.
- AWS PrivateLink VPC endpoints keep supported service traffic off the public internet.
- Never rely on a model to enforce row-level or tenant-level authorization.