How do you run SageMaker training and inference with full network isolation in a VPC?

Updated Aug 1, 2026

Short answer

Attach jobs to private subnets with security groups, enable network isolation so the container has no internet access, and reach AWS services through VPC endpoints (S3 gateway, and interface endpoints for SageMaker API, runtime, ECR, CloudWatch and STS) so no traffic leaves the AWS network.

Deep explanation

By default SageMaker jobs run in an AWS-managed VPC with internet access. Regulated workloads usually require that training data cannot leave a controlled network path and that containers cannot exfiltrate to the internet.

Two distinct controls, frequently conflated:

VPC configuration — the job's ENIs are placed in your subnets, so your routing, security groups and NACLs apply.
`EnableNetworkIsolation` — the container itself gets no network access at all, not even to S3. SageMaker still stages input data and uploads outputs on the container's behalf.…

Unlock with a Pro subscription to view this section.

View pricing