MMJailBench - Sensitive Regulated Advice: leaderboard
Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; Sensitive Regulated Advice domain, 30 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 | 3.56 |
| 2 | Claude Sonnet 4.5 | 11.22 |
| 3 | Qwen 3 VL 8B | 14.39 |
| 4 | InternVL3-8B | 27.39 |
| 5 | Kimi K2.5 | 30.72 |
| 6 | Qwen 2.5 VL 7B Instruct | 33.44 |
| 7 | Gemini 3 Flash | 36.61 |
| 8 | Ministral 3 8B | 38.06 |
| 9 | Grok 4 Fast | 40.33 |
| 10 | Step3 VL 10B | 42.06 |
| 11 | Seed 2.0 Lite | 46.44 |
| 12 | Gemma 3 12B | 51.39 |
| 13 | GLM-4.6V | 59.89 |
Interactive version: theaggregate.ai/benchmark?slug=mmjailbench-sensitive-regulated-advice · How It Works · Data refreshed daily, snapshot 2026-09-26.