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
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 0.29 |
| 2 | Claude Sonnet 4.5 | 0.88 |
| 3 | Qwen 3 VL 8B | 4.71 |
| 4 | Kimi K2.5 | 12.16 |
| 5 | Gemini 3 Flash | 12.35 |
| 6 | Seed 2.0 Lite | 16.18 |
| 7 | Step3 VL 10B | 27.06 |
| 8 | Ministral 3 8B | 30.59 |
| 9 | InternVL3-8B | 30.69 |
| 10 | Grok 4 Fast | 32.84 |
| 11 | Gemma 3 12B | 37.06 |
| 12 | Qwen 2.5 VL 7B Instruct | 50.78 |
| 13 | GLM-4.6V | 53.04 |
Interactive version: theaggregate.ai/benchmark?slug=mmjailbench-sexual-content · How It Works · Data refreshed daily, snapshot 2026-09-26.