Securing Toolbox with Model Armor

Mon, 01 Jan 0001 00:00:00 +0000

About

Google Cloud Model Armor is an LLM-agnostic service that screens prompts and responses to defend AI applications against prompt injection, jailbreaks, and sensitive data leakage. Pairing it with MCP Toolbox lets you screen both the prompts your users send and the responses your agent returns, including any sensitive data pulled from your tools, without trusting the model to police itself.

Model Armor screens traffic in two directions:

Ingress (incoming): Every input the model receives is screened before the model acts on it — the user’s prompt, and any data your tools return as it flows back in. This catches prompt injection and jailbreak attempts.
Egress (outgoing): Every response the model produces is screened before it returns to the user. This catches sensitive data leakage and harmful content.

sequenceDiagram
 actor User
 participant MA as Model Armor
 participant Agent as Agent / LLM
 participant Tool

 User->>MA: prompt
 Note over MA: Ingress: screen input
 MA->>Agent: prompt

 Agent->>Tool: tool call
 Tool->>MA: tool data
 Note over MA: Ingress: screen input
 MA->>Agent: tool data

 Agent->>MA: response
 Note over MA: Egress: screen output
 MA->>User: response

Note

Security on MCP Toolbox for Databases

Securing Toolbox with Model Armor

About