Model Armor integrates with Google Cloud networking services to provide inline security for AI applications at the network level. This integration allows Model Armor to inspect and filter prompts and responses without requiring modifications to the application code itself.
The mechanism for this integration is through Service Extensions . Service Extensions enable you to plug in custom logic, like Model Armor's security checks, into the data path of various Google Cloud networking services.
Before you begin
Ensure that you have met all the prerequisites .
Integration points
Model Armor integrates with the following networking services.
| Networking service | Description | Reference |
|---|---|---|
| Use Model Armor with Service Extensions to filter traffic passing through Layer 7 load balancers. This is useful for protecting web applications and APIs that leverage LLMs. | Configure a traffic extension to call the Model Armor service | |
| Use Model Armor with Service Extensions associated with the GKE Inference Gateway, for containerized AI applications or LLMs running on Google Kubernetes Engine, to inspect and enforce policies on traffic entering or leaving the GKE clusters. | Configure AI security and safety checks using GKE Inference Gateway | |
| Use Model Armor with Secure Web Proxy to inspect and secure egress traffic from your VPC to AI applications, MCP servers, or LLMs. | Configure an extension for Secure Web Proxy |
How it works
Service Extensions enable Google Cloud networking products to call out to Model Armor during data transiting.
- When traffic to or from your AI applications, MCP servers, or models passes through a configured network service, for example, a load balancer, Service Extensions forward the request or response content to Model Armor for inspection.
- Model Armor applies a set of filters based on your preconfigured template settings. These filters can identify and block prompt injections, jailbreak detection attempts, sensitive data leakage, malicious URLs, and harmful or inappropriate content, such as hate speech and harassment.
- Based on the scan results and your policy, Model Armor instructs the networking service to either allow, block, or modify the traffic, ensuring only safe and compliant interactions reach or leave your AI applications, MCP servers, or models.
