Securing AI Services I: Access and Data Privacy
Explore how to secure AWS AI services by configuring access permissions, employing IAM and service control policies, and protecting data privacy with private VPC endpoints and encryption. Understand Amazon Bedrock's model access, data handling practices, and how to control AI service usage in your organization.
We'll cover the following...
An application that sends prompts to a foundation model hands customer text to another service, so two questions need answers before it goes live. Who is allowed to use which model? Where does the data travel, and what happens to it afterward? A developer answers the first with model access settings and IAM policies, and the second with private network paths and the service’s data-use controls.
This lesson covers both for Amazon Bedrock, and then the opt-out setting that applies to other AWS AI services. The call itself, the bedrock-runtime client, and the execution role were covered in the lesson on calling Amazon Bedrock from code. Here, the focus is on restricting and protecting them.
The diagram below shows the controls on the path from an application to a model.
How access to a model is granted
Access to a model is a combination of an account-level step and an IAM permission. In commercial Regions, access to most serverless foundation models is enabled by default, and the account is subscribed the first time a model is invoked.
There is no request to submit for those models, and the old Model access page applies only to AWS GovCloud (US).
Three prerequisites decide whether that first invocation succeeds:
AWS Marketplace permissions: For third-party models sold through AWS Marketplace, the identity needs
aws-marketplace:Subscribe,aws-marketplace:Unsubscribe, andaws-marketplace:ViewSubscriptions. They are needed only the first time a model is used in an account. After someone with these permissions enables the model, other identities can invoke it without them.Anthropic first-time use case: Anthropic models also need a one-time use case form, submitted with
PutUseCaseForModelAccess, once per account or once in the organization’s management account (accounts in the organization inherit it).A valid payment method: The account must have a payment method that can be used for AWS Marketplace purchases.
Models from providers such as Amazon, Meta, and Mistral AI are not sold through AWS Marketplace, so there is no Marketplace permission to scope for them.
If prerequisites are missing, the subscription fails and calls return an AccessDeniedException. After the missing permission is granted, it can take a couple of minutes for the ...