MLA-C01 Practice Question 10: Training Network Isolation VPC
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 1 · question 10 of 10
A healthcare company trains models on de-identified patient records stored in Amazon S3. The training container is built from a public base image and installs third-party Python packages, and the security team does not fully trust it. Their requirements are that the training container must be unable to make any outbound network call, and that the training data must never traverse the public internet on its way into or out of the job.
Which configuration satisfies both requirements?
Show answer and explanation
Answer:
The two requirements need two different controls, and only this option applies both. Network isolation stops the container itself from making any outbound call to any service, including Amazon S3, and withholds AWS credentials from the container runtime; that is the control that contains an untrusted image. SageMaker AI still performs the S3 download of training data and the upload of model artifacts outside the container, using the execution role. Adding VpcConfig with private subnets and a gateway VPC endpoint for S3 routes that staging traffic through the customer VPC and over the AWS network rather than the public internet. AWS documents that network isolation and VPC configuration can be combined in exactly this way.
Option A addresses only the second requirement. Removing the NAT gateway blocks internet egress, but the container still has AWS credentials and network reachability inside the VPC, so a malicious package could still call any service exposed through a VPC endpoint or any host in the VPC.
Option B addresses only the first requirement. Network isolation genuinely prevents the container from calling out, but with default networking SageMaker AI performs the S3 data staging over the public AWS service endpoint rather than through the company's VPC.
Option D controls who may read the bucket, not what the container may do. A bucket policy scoped to a VPC endpoint would in fact break a job that is not running in that VPC, and it does nothing to prevent an untrusted library from exfiltrating data to an attacker-controlled host.
Reference: Run Training and Inference Containers in Internet-Free Mode