Security, Compliance, and Governance for AI Solutions
Shared responsibility
AWS secures the cloud: physical facilities, hardware, hypervisor, and the managed service itself. You secure what you put in the cloud: data classification, identity and permissions, configuration, and how prompts and outputs are handled. Managed AI services do not transfer your obligations around data governance to AWS.
Identity and data protection
Apply least privilege with IAM roles and policies, and use service control policies to set guardrails across accounts. Encrypt in transit with TLS and at rest with AWS KMS. Keep traffic private using VPC endpoints and AWS PrivateLink so model calls do not traverse the public internet. Store credentials in Secrets Manager rather than in code.
Foundation model data privacy
A frequently tested fact: with Amazon Bedrock, customer prompts and completions are not used to train the underlying base models and are not shared with third-party model providers, and data is processed within the selected region. You still control who can invoke which model through IAM.
Detection, audit, and governance
CloudTrail records API activity for audit, CloudWatch collects metrics and logs, AWS Config tracks configuration compliance, GuardDuty detects threats, Macie discovers sensitive data in Amazon S3, and AWS Artifact supplies compliance reports. For custom models, SageMaker Model Monitor detects drift and the model registry supports versioning and approval workflows.
Traps
- Encryption by default does not remove your duties; data classification, IAM, and prompt handling remain yours.
- CloudTrail records API calls, not the factual quality of model output.
- Region selection drives data residency; assuming a global service preserves residency is a common error.
Worked example and rule of thumb
For any control, ask: is this about the model and infrastructure, or about my data and configuration? Model weights and the managed runtime belong to AWS; prompts, identities, permissions, and monitoring belong to you.