Securing Azure OpenAI Access from On-Premises with Application Gateway: A Customer Success Story
vnamani shares a detailed case study on securing Azure OpenAI Service for an on-premises application using Application Gateway, focusing on network security and governance.
Securing Azure OpenAI Access from On-Premises with Application Gateway: A Customer Success Story
Introduction
A large enterprise customer needed to build a generative AI application hosted on-premises while leveraging Azure OpenAI Service. The primary challenge was ensuring secure and governed connectivity to Azure-based large language models (LLMs) without private network connectivity or a full Azure landing zone.
This walkthrough demonstrates how they utilized Azure Application Gateway as a reverse proxy—enabling secure, policy-driven access to Azure OpenAI and related services under strict security constraints.
Customer Landscape and Challenges
The customer’s environment was missing several core network components:
- No private network connectivity: No Site-to-Site VPN, ExpressRoute, or mature cloud network topology
- No Enterprise Scale Landing Zone (ESLZ): Lacked virtual WAN, hub-spoke, private DNS zones, DNS resolvers, API Management, and firewalls
- Strict security requirements: Public access to Azure OpenAI was not allowed
Key security objectives:
- Restrict access to specific on-prem IP CIDR ranges
- Limit permissible network ports
- Implement SSL termination and a Web Application Firewall (WAF)
- Use customer-managed SSL certificates
Proposed Solution Architecture
To bridge these gaps, the team implemented a secure Azure-centric architecture using:
Core Azure Services
- Application Gateway: Provided Layer 7 reverse proxy, SSL termination, and WAF capabilities
- Public IP: Enabled internet-based communication from authorized on-prem addresses
- Virtual Network: Allowed segmentation and control of network traffic
- Network Security Groups (NSG): Managed inbound/outbound rules, only opening required HTTP(S) ports and whitelisting specific IP ranges
- Azure OpenAI: Served as the destination for LLM queries and responses
Implementation Details
NSG Configuration
- Inbound rules: Allowed only traffic from customer’s IP CIDR ranges over appropriate HTTP(S) ports
- Outbound rules: Targeted AzureCloud endpoints, acknowledging lack of direct Azure OpenAI service tags
Application Gateway Deployment
- SSL Certificate: Provided by the customer’s on-premise Certificate Authority for HTTPS termination
- Traffic flow:
- Decrypt incoming traffic at the Application Gateway
- Inspect using the WAF for threats
- Re-encrypt outgoing traffic with Azure-trusted CA
- Override backend hostname as needed
- Health probe: Customized to recognize an expected 404 response from Azure OpenAI (since no native health endpoint exists)
Azure OpenAI Configuration
- IP firewall: Only allowed incoming traffic from the Application Gateway subnet, reducing exposure
Outcome
This solution enabled the customer to:
- Secure Azure OpenAI Service using well-understood Azure and networking primitives
- Operate without a complete Azure landing zone or private connectivity
- Meet enterprise-grade security and governance standards using layered defense (NSG, WAF, SSL, IP firewall)
The architecture can serve as a reference for others facing similar hybrid connectivity and security challenges when deploying Azure OpenAI in regulated or constrained network environments.
Author: vnamani
This post appeared first on “Microsoft Tech Community”. Read the entire article here