OpenAI CSAM Classifier - Held-out CSAM AUROC: leaderboard

Publisher evaluation of one released CLIP+MLP image classifier. The two output signals belong to the same checkpoint. Held-out images were randomly split and perceptually deduplicated; related images can remain across splits. Native fractions, higher is better. The held-out set contains 46,849 images; the GenAI positives are a subset of CSAM positives. AUROC and AUPRC integrate over thresholds. Metric: AUROC (0–1). Source: https://deploymentsafety.openai.com/csam-image-classifier/csam-image-classifier.pdf (table 3, page 8).

Metric: AUROC (0–1). Source: deploymentsafety.openai.com. 1 model tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1CSAM Image Classifier0.998

Interactive version: theaggregate.ai/benchmark?slug=openai-csam-classifier-held-out-csam-auroc · How It Works · Data refreshed daily, snapshot 2026-11-17.