MMJailBench - Academic Framing: leaderboard
Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; academic-analysis framing, all 272 intents, visual contexts and carriers). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 0.74 |
| 2 | Claude Sonnet 4.5 | 5.4 |
| 3 | Qwen 3 VL 8B | 25.37 |
| 4 | Gemini 3 Flash | 34.96 |
| 5 | Kimi K2.5 | 59.45 |
| 6 | Seed 2.0 Lite | 63.27 |
| 7 | Step3 VL 10B | 65.11 |
| 8 | Grok 4 Fast | 66.4 |
| 9 | InternVL3-8B | 74.41 |
| 10 | Ministral 3 8B | 82.72 |
| 11 | GLM-4.6V | 84.67 |
| 12 | Gemma 3 12B | 86.4 |
| 13 | Qwen 2.5 VL 7B Instruct | 91.43 |
Interactive version: theaggregate.ai/benchmark?slug=mmjailbench-academic-framing · How It Works · Data refreshed daily, snapshot 2026-09-26.