Diagram

A sequence diagram showing an app sending prompt plus safety settings to gemini-2.5-flash, the model returning candidates with four safety ratings, and the app blocking on harassment and showing a fallback.
Automated Defenses With Safety Settings
Automated defenses with safety settings are controls in Vertex AI that evaluate every generation and block unsafe output without human review. Data flows from the application to the model, in this case the Gemini model gemini-2.5-flash, as a package containing the user prompt, an optional system instruction, and a list of safety settings. The model returns candidates that contain generated text plus safety ratings, and each rating reports its category, whether it was blocked, and its probability and severity scores. When a rating meets or exceeds its configured threshold, generation is blocked before it reaches the user.
Domain 2 of 3
35 practice questions
Practice this DomainLast updated