MMJailBench - Structured Framing: leaderboard
Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; structured-completion framing, all 272 intents, visual contexts and carriers). Source: arxiv.org. Saturation forecast: Around December 2026. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 8.82 |
| 2 | Kimi K2.5 | 20.96 |
| 3 | Qwen 3 VL 8B | 29.12 |
| 4 | Claude Sonnet 4.5 | 32.13 |
| 5 | Seed 2.0 Lite | 45.77 |
| 6 | Gemini 3 Flash | 50.88 |
| 7 | InternVL3-8B | 57.13 |
| 8 | Grok 4 Fast | 74.15 |
| 9 | Ministral 3 8B | 76.36 |
| 10 | Qwen 2.5 VL 7B Instruct | 84.67 |
| 11 | Step3 VL 10B | 87.68 |
| 12 | Gemma 3 12B | 92.21 |
| 13 | GLM-4.6V | 93.86 |
Interactive version: theaggregate.ai/benchmark?slug=mmjailbench-structured-framing · How It Works · Data refreshed daily, snapshot 2026-09-26.