MMJailBench - Hate & Harassment: leaderboard

Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; Hate & Harassment domain, 30 harmful intents across all framings, visual contexts and carriers). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.

Top models

#ModelScore
1GPT-52.5
2Claude Sonnet 4.57
3Qwen 3 VL 8B12.06
4Gemini 3 Flash28.72
5Kimi K2.534
6InternVL3-8B42.33
7Ministral 3 8B43.11
8Seed 2.0 Lite47.5
9Step3 VL 10B49.17
10Grok 4 Fast49.33
11Qwen 2.5 VL 7B Instruct54.39
12Gemma 3 12B62.67
13GLM-4.6V74.94

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