MMJailBench - Sexual Content: leaderboard

Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; Sexual Content domain, 17 harmful intents across all framings, visual contexts and carriers). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.

Top models

#ModelScore
1GPT-50.29
2Claude Sonnet 4.50.88
3Qwen 3 VL 8B4.71
4Kimi K2.512.16
5Gemini 3 Flash12.35
6Seed 2.0 Lite16.18
7Step3 VL 10B27.06
8Ministral 3 8B30.59
9InternVL3-8B30.69
10Grok 4 Fast32.84
11Gemma 3 12B37.06
12Qwen 2.5 VL 7B Instruct50.78
13GLM-4.6V53.04

Interactive version: theaggregate.ai/benchmark?slug=mmjailbench-sexual-content · How It Works · Data refreshed daily, snapshot 2026-09-26.