Netskope AI Guardrails on Demand

AWS deployment walkthrough using CloudFormation

This guide walks you through deploying the Netskope AI Guardrails on Demand appliance into AWS with a single CloudFormation stack, registering it with your Netskope tenant, and testing it. AI Guardrails on Demand ships as a Netskope Virtual Private Edge (VPE) virtual appliance. It scans AI prompts and responses through an API, so an AI gateway or agent you already run can call it and enforce the verdict.

Beta. AI Guardrails on Demand is a Beta feature. Ask Netskope Support or your account team to enable it for your tenant and to share the VPE appliance AMI with your AWS account.

The Stack

netskope-guardrails-standalone

What the stack builds

One aws cloudformation deploy creates everything below. No existing network is required.

• A new VPC with an internet gateway, two public and two private /24 subnets across two Availability Zones, one NAT gateway, route tables and an S3 gateway endpoint.
• The VPE appliance in an Auto Scaling group (one instance), in the private subnets. It reaches the Netskope management plane outbound through the NAT gateway.
• An internal Application Load Balancer that serves HTTPS on 443 with a self-signed certificate and forwards to the appliance API on port 8080.
• A Route 53 private hosted zone that resolves guardrails.aigw.internal to the load balancer.
• Security groups, the self-signed certificate (stored in ACM and Systems Manager Parameter Store) and a small Lambda that generates it.

Infrastructure only. The stack builds the AWS side. Enrolling the appliance in your tenant and attaching the Netskope templates are manual steps, covered in sections 6 to 10.

Before you start

PowerShell tip. Commands in this guide are written for PowerShell, on one line each. If you copy a multi-line command from elsewhere that ends lines with a backslash, that is bash syntax and PowerShell will reject it. Put it on one line, or end each line with a backtick instead.

1. Check your AWS credentials

On your computer, confirm which AWS account you are about to deploy into. The command changes nothing; it shows the account and identity your credentials resolve to.

aws sts get-caller-identity

If it returns an error, sign in again (for example aws sso login) before you continue.

2. Get the repository

Clone the repository and change into it. Run the remaining commands from this folder.

git clone https://github.com/sandiegojenkins/netskope-guardrails-standalone.git

cd netskope-guardrails-standalone

Credentials check, then clone the repository.

3. Configure the deployment

Copy the example settings file to .env and edit it. The file lives in the repository folder, next to deploy.ps1, and is excluded from git, so it never gets committed.

Copy-Item .env.example .env

Open .env in a text editor and set the values below. Some defaults are already filled in.

Important. Save the file as .env (no .txt extension). If .env is in a different folder than deploy.ps1, the script will not find it.

Example .env. The VPC_CIDR line is still commented out, so the default is used.

4. Deploy the stack

Run the deploy script from the repository folder.

./deploy.ps1

The script reads .env, runs aws cloudformation deploy, waits for the stack to finish and then prints the stack outputs. The NAT gateway and load balancer take the longest to create, so allow several minutes.

A successful deployment ends with the stack outputs table.

Keep these outputs handy; later steps use them.

If the deploy fails. CloudFormation rolls back and leaves the stack in ROLLBACK_COMPLETE. Read the cause with aws cloudformation describe-stack-events --stack-name <stack-name> --region <region>, fix it, then delete the failed stack before deploying again: aws cloudformation delete-stack --stack-name <stack-name> --region <region>.

5. Confirm the appliance is running

The stack launches the appliance through an Auto Scaling group. Check that the instance is InService, and note its instance ID. Replace the group name with your GuardrailsAsgName output and the region with yours.

aws autoscaling describe-auto-scaling-groups --auto-scaling-group-names <GuardrailsAsgName> --query "AutoScalingGroups[0].Instances[*].[InstanceId,LifecycleState,HealthStatus]" --region <region> --output table

One instance, InService and Healthy.

Why the group does not replace the appliance. The group uses an EC2 health check, not a load balancer health check. Until you finish activation, the appliance API may not answer and the load balancer may report the target unhealthy. That is expected, and the instance is not terminated.

6. Connect to the appliance CLI

The appliance sits in a private subnet and the stack has no bastion host. If you do not already have VPN, peering or Transit Gateway access into the VPC, use an EC2 Instance Connect Endpoint to reach it. The endpoint is free and is created outside the stack (see Teardown).

1. Create the endpoint in one of the private subnets (from the PrivateSubnetIds output). It can take a few minutes to become ready.

aws ec2 create-instance-connect-endpoint --subnet-id <private-subnet-id> --region <region>

2. Open a tunnel to the instance. Leave this window running.

aws ec2-instance-connect open-tunnel --instance-id <instance-id> --local-port 2222 --region <region>

3. Windows requires the key file to be private to your account. Run this once for your .pem file.

icacls .\<key-name>.pem /inheritance:r

icacls .\<key-name>.pem /grant:r "$($env:USERNAME):R"

4. In a second PowerShell window, connect as nsadmin.

ssh -i .\<key-name>.pem -p 2222 nsadmin@localhost

You should see the Netskope VPE banner and a CLI prompt such as ip-10-5-3-40>.

Good to know. The username is nsadmin, not root. The VPE appeared to accept only one SSH session at a time, so type exit before opening another. If you replace the instance later and SSH warns that the host key changed, run ssh-keygen -R "[localhost]:2222" and connect again.

7. Register the appliance with your tenant

Create a registration token in the Netskope UI

1. Go to Settings > Security Cloud Platform > On-Premises Infrastructure.
2. Select the Next-Gen tab and click Registration Tokens.
3. In the Registration Tokens window, click Create Token. Choose how many tokens (one is enough), then copy or download the token.

Enter the token in the appliance CLI

In the SSH session, run these commands in order. Replace <vpc-resolver-ip> with the VPC base address plus 2 (for the default 10.5.0.0/16 that is 10.5.0.2), and <token> with your token.

configure

set dns primary <vpc-resolver-ip>

set system registrationkey <token>

save

exit

status tethering

You must run save. Without save the registration key is not applied. The node stays not_registered (only an identifier appears) no matter how long you wait. The DNS line is an explicit setting; the node normally picks up the VPC resolver through DHCP, which you can check with show dns.

Registration can take up to 20 minutes. When it completes, status tethering shows "status": "tethered" with your tenant_url and a serial, and the node appears on the Next-Gen page with its hostname and serial number. Treat the registration token like a password, and do not paste it into tickets or screenshots.

8. Create and apply the templates

A registered node does nothing until it has a platform template and a service template. Create both, then attach them to the node. Both must be attached before the AI Guardrails API on the appliance answers.

Create the platform template

1. Go to Settings > Security Cloud Platform > On-Premises Infrastructure and select the Next-Gen tab.
2. Click Manage Template, then Platform Template, then New Template.
3. Enter a VPE Platform Template Name and an NTP Server (both required). The example uses 0.us.pool.ntp.org. The description and upgrade timing are optional.
4. Click Save.

Platform template: a name and an NTP server are required.

Create the AI Guardrails service template

1. On the Next-Gen tab, click Manage Template, then Service Template, then New Template.
2. Enter a Service template name.
3. Under Services, select AI Guardrails (not SWG).
4. Under General Settings, choose the listener option. The example template uses HTTPS (Port 443); the certificate and private key are uploaded through the appliance CLI (see section 10).
5. Under GPU based Detection Settings, leave GPU based AI Guardrails VM off unless you run the optional GPU backend. That backend is a separate VM and is not part of this stack.
6. Under Store Prompts and Responses Settings, decide whether to turn on Store Matched Content in Netskope. When it is on, matched text is stored for Skope IT alerts and incidents, and the API echoes the evaluated text back in its response.
7. Click Save.

Service template with AI Guardrails selected. The load balancer in this stack forwards to the appliance API on port 8080.

Attach both templates to the node

1. On the Next-Gen page, find your node and click Appliance Setup in the Services column.
2. Choose your platform template from the VPE Platform Template dropdown and your service template from the VPE Service Template dropdown.
3. Click Save.

Appliance Setup: choose the platform and service templates, then Save.

9. Check the load balancer target

Once both templates are attached, the appliance serves the API. Give it a few minutes, then check the load balancer target health. The target should show healthy.

To find the target group ARN, run aws elbv2 describe-target-groups --region <region> --query "TargetGroups[].[TargetGroupName,TargetGroupArn]" --output table.

10. Generate the dataplane certificate

In the appliance CLI, generate the self-signed dataplane certificate. Use any common name and your own email address.

A successful run prints successfully generated self signed ca.

11. Test the service

The load balancer is internal, so run the test from inside the VPC. The simplest way is an AWS CloudShell VPC environment, which gives you a shell in your VPC that resolves the private name guardrails.aigw.internal and is already allowed by the load balancer security group. (An Instance Connect Endpoint cannot carry this traffic because it only tunnels ports 22 and 3389.)

Create the CloudShell VPC environment

1. In the AWS console, switch to your deployment region and open CloudShell.
2. Choose Actions, then Create VPC environment (global actions section).
3. Enter a name, then pick the VPC (the VpcId output), one of the private subnets (PrivateSubnetIds output) and the VPC's default security group. Create the environment and wait for it to start.

Get your profile ID

The test needs the ID of an AI Guardrails profile in your tenant. In the Netskope UI go to Policies > Profiles > AI Guardrails > Profiles, click the three dots on your profile and choose Copy Profile ID.

Send a test prompt

In the CloudShell VPC environment, replace <profile-id> and run the command. -k is needed because the load balancer certificate is self-signed.

A prompt that breaks a rule returns a Match verdict. A harmless prompt returns Not Match. The first example below was a real run on this deployment (profile ID shown as sent):

And a harmless prompt on the same deployment:

Empty reply. If a request prints nothing, run it again, and add -v to see the HTTP status. One empty reply was seen during testing and the identical retry returned normally.