Model Armor is a Google Cloud service designed to enhance the security and safety of your AI applications, particularly those using Large Language Models (LLMs). It works by inspecting the prompts sent to your models and the responses generated by them, helping you mitigate risks and enforce responsible AI practices.
Configure templates
Define how Model Armor should screen content by creating and using Model Armor templates. A template is a reusable configuration set where you specify which filters to enable, the confidence levels for the filters, and the enforcement type for each filter. For more information, see Create and manage templates.
Configure floor settings
To ensure a baseline level of protection, security administrators can configure floor settings at the organization, folder, or project level. These settings mandate minimum filter requirements that all Model Armor templates created within that scope must adhere to, helping to prevent overly permissive configurations. For more information, see Configure floor settings.
Sanitize prompts and responses
When a user submits a prompt to your application, your application first sends this prompt to Model Armor. Model Armor processes the prompt through the enabled filters in the template and returns a response indicating whether any policy violations were found and detailed results from each filter. Your application logic then decides what to do next.
When an LLM generates a response, before displaying this response to the user, your application sends it to Model Armor. Model Armor screens the LLM output using the filter configurations defined in the template and returns the analysis results. Your application then decides whether to show the response to the user, potentially blocking it if violations are found.
For more information, see Sanitize prompts and responses.