Turkish Injection Guards - False Positive Rate: leaderboard

Turkish Injection Guards - False Positive Rate. This Turkish prompt-injection benchmark compares three released guard classifiers on 317 attacks from five sources and 3,055 benign examples. Each classifier uses a 0.5 positive-class threshold and a maximum of 512 input tokens. The source reports each attack subset’s recall, their unweighted mean, positive-class F1, and the false-positive rate on benign examples. Proventra’s Berat result has a disclosed training-source overlap. The author’s v1, v2, and ensemble summary rows combine source-held-out training artifacts and are not mapped to one final checkpoint. These are measured classifier capabilities; ingestion alone does not decide a downstream global-fit policy. Reported metric: False Positive Rate (%).

Metric: False Positive Rate (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 3 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=turkish-injection-guards-false-positive-rate · How It Works · Data refreshed daily, snapshot 2026-10-25.