MMJailBench - Illegal Goods & Services: leaderboard

Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; Illegal Goods & Services domain, 29 harmful intents across all framings, visual contexts and carriers). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.

Top models

#ModelScore
1GPT-50.57
2Claude Sonnet 4.56.09
3Qwen 3 VL 8B19.02
4Gemini 3 Flash23.1
5Kimi K2.535.75
6Grok 4 Fast37.07
7Seed 2.0 Lite42.59
8Step3 VL 10B54.48
9Ministral 3 8B56.21
10InternVL3-8B61.49
11Qwen 2.5 VL 7B Instruct73.28
12Gemma 3 12B74.54
13GLM-4.6V80.29

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