Prerequisites and permissions

Prerequisites

When you submit an Advanced Prompt Optimization job, Amazon Bedrock accesses resources in your account on your behalf using the permissions of your calling identity. The IAM user or role that you use to call the API or access AWS console must have permissions to:

Required IAM permissions

API permissions

Each Advanced Prompt Optimization API requires a specific IAM action on the job resource:

Model invocation permissions

When you create a job, Advanced Prompt Optimization invokes the target model(s) on your behalf using your caller credentials (Forward Access Session). Your IAM identity must have permission to invoke the specified models:

S3 bucket permissions

Your calling identity must have read access to the input dataset and write access to the output location. The S3 bucket must be in the same Region as the job.

{
    "Effect": "Allow",
    "Action": [
        "s3:GetObject",
        "s3:PutObject"
    ],
    "Resource": "arn:aws:s3:::DOC-EXAMPLE-BUCKET/*"
},
{
    "Effect": "Allow",
    "Action": "s3:ListBucket",
    "Resource": "arn:aws:s3:::DOC-EXAMPLE-BUCKET"
}

If your S3 bucket uses CMK encryption, see the Encryption with AWS KMS section for the required KMS permissions.

Encryption with AWS KMS (optional)

You can use AWS KMS customer-managed keys (CMKs) with Advanced Prompt Optimization in two ways:

You can use the same KMS key for both, or use separate keys.

The required KMS actions depend on how you use KMS:

Single KMS key

If you use the same key for both output encryption and S3 bucket encryption:

{
    "Effect": "Allow",
    "Action": [
        "kms:Encrypt",
        "kms:Decrypt",
        "kms:GenerateDataKey",
        "kms:DescribeKey"
    ],
    "Resource": "arn:aws:kms:us-west-2:123456789012:key/your-kms-key-id"
}

Separate KMS keys for S3 input and job output encryption

If your S3 input bucket uses one KMS key and you want to encrypt job output with a different KMS key:

{
    "Sid": "KMSDecryptS3Input",
    "Effect": "Allow",
    "Action": [
        "kms:Decrypt"
    ],
    "Resource": "arn:aws:kms:us-west-2:123456789012:key/s3-bucket-kms-key-id"
},
{
    "Sid": "KMSEncryptJobOutput",
    "Effect": "Allow",
    "Action": [
        "kms:Encrypt",
        "kms:GenerateDataKey",
        "kms:DescribeKey"
    ],
    "Resource": "arn:aws:kms:us-west-2:123456789012:key/output-encryption-kms-key-id"
}

Custom evaluation Lambda permissions (optional)

If you use a custom Lambda function to evaluate prompt quality, your calling identity must have permission to invoke it:

{
    "Effect": "Allow",
    "Action": ["lambda:InvokeFunction","lambda:GetFunction"],
    "Resource": "arn:aws:lambda:us-west-2:123456789012:function:my-evaluation-metric"
}

Sample IAM policies

Identity-based policy: Allow AdvPO job management and model invocation

Attach this policy to the IAM user or role that will create and manage AdvPO jobs.

{
    "Version": "2012-10-17", 
    "Statement": [
        {
            "Sid": "CreateAndManageAdvPOJobs",
            "Effect": "Allow",
            "Action": [
                "bedrock:CreateAdvancedPromptOptimizationJob",
                "bedrock:GetAdvancedPromptOptimizationJob",
                "bedrock:ListAdvancedPromptOptimizationJobs",
                "bedrock:StopAdvancedPromptOptimizationJob",
                "bedrock:BatchDeleteAdvancedPromptOptimizationJob"
            ],
            "Resource": "arn:aws:bedrock:us-east-1:123456789012:advanced-prompt-optimization-job/*"
        },
        {
            "Sid": "InvokeModelsForOptimization",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:*::foundation-model/*",
                "arn:aws:bedrock:*:123456789012:inference-profile/*"
            ]
        },
        {
            "Sid": "S3AccessForInputAndOutput",
            "Effect": "Allow",
            "Action": [
                "s3:GetObject",
                "s3:PutObject"
            ],
            "Resource": "arn:aws:s3:::my-advpo-bucket/*"
        },
        {
            "Sid": "S3ListBucket",
            "Effect": "Allow",
            "Action": "s3:ListBucket",
            "Resource": "arn:aws:s3:::my-advpo-bucket"
        }
    ]
}

Identity-based policy: Read-only access to AdvPO jobs

Use this policy for users who only need to view job status and results.

{
    "Version": "2012-10-17", 
    "Statement": [
        {
            "Sid": "ViewAdvPOJobs",
            "Effect": "Allow",
            "Action": [
                "bedrock:GetAdvancedPromptOptimizationJob",
                "bedrock:ListAdvancedPromptOptimizationJobs"
            ],
            "Resource": "arn:aws:bedrock:us-east-1:123456789012:advanced-prompt-optimization-job/*"
        },
        {
            "Sid": "ReadOutputFromS3",
            "Effect": "Allow",
            "Action": "s3:GetObject",
            "Resource": "arn:aws:s3:::my-advpo-bucket/output/*"
        }
    ]
}

Identity-based policy: Restrict to specific models

Use this policy to allow AdvPO only with specific models (for example, only Claude Sonnet).

{
    "Version": "2012-10-17", 
    "Statement": [
        {
            "Sid": "AdvPOJobManagement",
            "Effect": "Allow",
            "Action": [
                "bedrock:CreateAdvancedPromptOptimizationJob",
                "bedrock:GetAdvancedPromptOptimizationJob",
                "bedrock:ListAdvancedPromptOptimizationJobs",
                "bedrock:StopAdvancedPromptOptimizationJob"
            ],
            "Resource": "arn:aws:bedrock:us-east-1:123456789012:advanced-prompt-optimization-job/*"
        },
        {
            "Sid": "InvokeOnlyClaudeSonnet",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:*::foundation-model/anthropic.claude-sonnet-*",
                "arn:aws:bedrock:*:123456789012:inference-profile/us.anthropic.claude-sonnet-*"
            ]
        },
        {
            "Sid": "S3Access",
            "Effect": "Allow",
            "Action": [
                "s3:GetObject",
                "s3:PutObject"
            ],
            "Resource": "arn:aws:s3:::my-advpo-bucket/*"
        }
    ]
}

Deny policy: Prevent AdvPO usage (SCP example)

Use as a Service Control Policy (SCP) to prevent AdvPO usage across an organization.

{
    "Version": "2012-10-17", 
    "Statement": [
        {
            "Sid": "DenyAdvPO",
            "Effect": "Deny",
            "Action": [
                "bedrock:CreateAdvancedPromptOptimizationJob",
                "bedrock:StopAdvancedPromptOptimizationJob",
                "bedrock:BatchDeleteAdvancedPromptOptimizationJob"
            ],
            "Resource": "*"
        }
    ]
}

Cross-region inference

The service may use cross-region inference (CRIS) for evaluation and prompt rewriting. For European regions, the service will use eu CRIS regions. For the following regions, it may use global CRIS: ap-south-1 (Mumbai), sa-east-1 (Sao Paulo), ap-northeast-2 (Seoul), ap-southeast-1 (Singapore).