Flags unsafe user prompts and unsafe model responses, with harm categories, using content-safety-guard-small (141M, multilingual, Apache-2.0), running entirely in your browser. The first run downloads about 270 MB, which is then cached. The base model is somewhat more accurate.
User prompt
Model response (optional: leave empty to check the prompt only)
The score is the model's probability that the prompt (or, with a response, the response) is unsafe; the default threshold is 0.5. Categories are shown when flagged, using per-category thresholds from the model repository. It follows a general-purpose policy (distilled from Qwen3Guard-Gen-8B) and makes mistakes: it over-flags some harmless requests that sound dangerous and misses some harmful ones. See the model card for measured accuracy.