Predefined groups of harmful content types (like harassment or hate speech) that a safety model is trained to detect and flag.