If an attacker has sufficient permissions to initiate these SageMaker resources and holds a broad iam:PassRole grant, they can pass a more privileged role - or a role from a different account - to the resource. The resulting SageMaker resource then executes code under the context of the assigned role, effectively allowing the attacker to perform actions that the original caller could not. This technique is particularly dangerous as it utilizes legitimate, expected functionality to bridge security boundaries. Monitoring for unusual associations between IAM principals and execution roles is essential for detecting potential credential abuse or unauthorized privilege escalation within AWS MLOps workflows.

Impact

Successful exploitation allows an attacker to achieve code execution as a more privileged IAM role, leading to unauthorized data access, persistence, or broader administrative control within the AWS account. This risk is especially critical in production MLOps environments where high-privilege service roles are frequently utilized for legitimate automated data processing.

Recommendation

• Implement the detection logic to identify the first time an IAM principal uses a specific roleArn for SageMaker resource creation within a 7-day window.
• Review the aws.cloudtrail.user_identity.arn of the principal performing the action and investigate the roleArn passed in aws.cloudtrail.request_parameters for unauthorized privilege escalation.
• Constrain iam:PassRole permissions using IAM policy conditions (such as iam:PassedToService) to ensure that principals can only pass narrowly scoped, approved execution roles to SageMaker.
• Rotate or restrict credentials for any principal identified using an unexpected or overly privileged execution role.