Configuring AI File Classification and Entity Detection for DLP On Demand

The DLP On Demand (DLPoD) appliance supports three AI-powered content inspection capabilities, referred to together as DLP-LLM.

You enable these capabilities from nsshell with two independent commands:

• request dlp-llm configure sets up AI File Classification.
• request dlp-llm-ner configure sets up NER and genomic detection.

The two commands use separate credentials and don’t affect each other. You can enable one without the other.

DLP-LLM inference always runs in AWS, regardless of where your DLPoD appliance is hosted. The appliance itself supports AWS, Azure, GCP, and on-premises hypervisors (Hyper-V, KVM, and VMware), but you need an AWS account to host the inference backends either way. The authentication method the appliance uses depends on where the appliance runs.

If your appliance runs on AWS EC2, an EC2 instance role is the default and recommended option for both capabilities. The role’s credentials rotate automatically, so you don’t manage or rotate any keys. If your appliance runs on Azure, GCP, or on-premises, it has no native AWS identity, so it must authenticate with a Bedrock long-term API key for AI File Classification and a JWT for NER.

Prerequisites

Before you configure DLP-LLM, confirm the following:

• Your DLPoD appliance is deployed and fully initialized, on a release that includes the dlp-llm service.
• Your appliance is tethered. The configurator encrypts credentials with the tenant transit key, which is only available after tethering.
• Your appliance has network egress to AWS Bedrock and, for NER, to your SageMaker endpoint. If your egress uses SSL inspection, see the troubleshooting section for the CA bundle option.
• You have an AWS account. Bedrock and the SageMaker NER endpoint always run in AWS, regardless of your appliance’s hosting platform.

For AI File Classification, you also need AWS credentials the appliance can use to call Bedrock. For NER, you also need a SageMaker real-time endpoint in your AWS account and credentials that let the appliance invoke it.

If you plan to set up the SageMaker endpoint yourself, you also need these local tools:

• AWS CLI v2
• Docker, with the daemon running
• crane, which the NER deployment script uses to push images in the manifest format SageMaker requires
• IAM permissions to create ECR repositories, IAM roles and policies, and SageMaker models and endpoints

Configuring AI File Classification

AI File Classification runs on AWS Bedrock. The IAM permissions you need depend on which model you select.

Setting Up AWS IAM for Bedrock

Two model families are available, and each uses a different IAM policy.

• google.gemma-4-31b and openai.gpt-5.6-luna run on Bedrock Mantle.
• us.amazon.nova-pro-v1:0 and us.amazon.nova-2-lite-v1:0 run on standard Bedrock.

For the Bedrock Mantle models, attach a policy that allows bedrock-mantle:CreateInference and bedrock-mantle:CallWithBearerToken, scoped to your account’s default Bedrock Mantle project.

For the standard Bedrock models, attach a policy that allows bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream. Cross-region model calls are checked against two resources, the inference profile and the underlying foundation model. Include both in the policy.

If your appliance runs on AWS EC2, you can attach an instance profile instead of using static keys. Create the role, attach the Bedrock policy, create an instance profile, and associate it with your running instance using the AWS CLI. This requires no restart of the instance.

If your appliance runs off AWS (Azure, GCP, or on-premises), generate a Bedrock long-term API key instead:

1. In the AWS console, go to Amazon Bedrock > API Keys.
2. Under Long-term API Keys, select Generate Long-Term API Key.
3. Specify an expiration.
4. Copy the key. You enter it as the API key credential in request dlp-llm configure.

Running request dlp-llm configure

From nsshell on the DLPoD appliance, run:

nsappliance> request dlp-llm configure

The command prompts for the following values. Defaults are shown in brackets.

1. Auth method: aws-keys (default) or api-key. Choose aws-keys if your appliance runs on AWS EC2, or api-key if it runs off AWS.
2. AWS credentials, matching the auth method you chose.
3. Region, the AWS region for Bedrock. Default is us-west-2.
4. Daily request limit, service.max_daily_requests. Enter 0 for no limit. Default is 100000.
5. Model, one of the four supported model IDs. Default is google.gemma-4-31b.

A model change takes effect the next time the dlp-llm service restarts, not on a configuration reload.

Configuring NER and Genomic Entity Detection

NER and genomic detection run against a SageMaker real-time endpoint in your own AWS account. You host the endpoint, and the appliance invokes it. Content sent for inspection never leaves your AWS environment.

Deploying the SageMaker Endpoint

Netskope provides a container image with the NER and genomic models built in, delivered as a compressed Docker image tarball or as a pull reference to a Netskope-hosted registry.

Confirm the following before you deploy:

Netskope provides a deployment script, deploy_ner_sagemaker.sh, that automates the full setup: it loads the image, pushes it to ECR, creates the SageMaker execution role if one doesn’t exist, creates the model, endpoint configuration, and endpoint, and validates both the PII and genomic model families with a test call. Run it with your tarball, AWS account ID, and region:

./deploy_ner_sagemaker.sh --tarball dlp-ner-sagemaker-<version>.tar.gz \
  --account <your-aws-account-id> --region <your-region>

To validate a running endpoint without redeploying, run the script with –validate-only.

Setting Up NER Authentication

You choose how the appliance authenticates to your SageMaker endpoint. Four credential types are available.

Whichever type you choose, grant only the minimum permission, sagemaker:InvokeEndpoint, scoped to your endpoint’s Amazon Resource Name (ARN).

If your appliance runs on AWS EC2 (Type D, recommended): attach an IAM role to the appliance’s EC2 instance with the invoke permission above. If the instance already has a role, add the permission to it instead of replacing the role, since an instance can hold only one role. When you run request dlp-llm-ner configure, leave the AWS credential fields blank. The appliance then uses the instance role automatically. If the dlp-llm service is already running, restart it so it picks up the role.

If your appliance runs off AWS (Type B, required): front your SageMaker endpoint with Amazon API Gateway and a Cognito authorizer, so the appliance holds a long-lived Cognito client secret instead of an AWS access key. At a high level:

1. Create a Cognito user pool, domain, resource server, and scope.
2. Create a machine-to-machine app client, which generates a client ID and client secret.
3. Create an IAM role that allows API Gateway to invoke your SageMaker endpoint.
4. Create an API Gateway REST API with a POST method, protected by the Cognito authorizer, that integrates with your SageMaker endpoint.
5. Deploy the API and note its URL.

You then enter the following values in request dlp-llm-ner configure:

Running request dlp-llm-ner configure

From nsshell on the DLPoD appliance, run:

nsappliance> request dlp-llm-ner configure

The command prompts for the following values. Defaults are shown in brackets.

1. Auth method: aws-keys (default), api-key, or jwt. Choose based on the credential type you set up above.
2. Credentials, matching the auth method you chose. For a permanent IAM key or EC2 instance role, this is an AWS access key ID and secret (leave both blank to use an instance role). For JWT, this is your client ID, client secret, token URL, API URL, and scope.
3. Region, the AWS region of your SageMaker endpoint. Default is us-west-2.
4. Daily request limit, service.max_daily_requests. Enter 0 for no limit. Default is 100000.

This command doesn’t prompt for a model selection. The NER model is part of your SageMaker endpoint, not a Bedrock model.

Verifying and Troubleshooting

To check the appliance’s overall health and confirm which services are running, use:

nsappliance> request status health
nsappliance> request status all

To confirm which authentication method DLP-LLM is using, check the dlp-llm service logs for a message that names the active method, such as an IAM permanent key, a temporary key, JWT, or an EC2 instance role.

The following table lists common issues.

Security Notes

Quick Command Reference

request dlp-llm configure        # AI File Classification (Bedrock)
request dlp-llm-ner configure    # NER and genomic detection (SageMaker)