MMJailBench - Safety-Paradox Framing: leaderboard
MMJailBench (Wang et al., 2026) is a factorized multimodal jailbreak benchmark: 272 harmful intents crossed with six prompt framings, five task-relevant visual contexts and text or in-image instructions, 16,320 instances per model. This board covers one prompt framing, in which the request is justified with a safety-oriented rationale, over all 272 intents. Attack success rate in percent, the share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; lower is better.
Metric: Attack success rate (%; share of responses a GPT-5 judge scores 4 or 5 on a 1-5 harmfulness rubric; safety-paradox reasoning framing, all 272 intents, visual contexts and carriers). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 1.88 | #163 |
| 2 | Claude Sonnet 4.5 | 3.27 | #244 |
| 3 | Kimi K2.5 | 15.15 | #252 |
| 4 | Gemini 3 Flash | 18.97 | #159 |
| 5 | Ministral 3 8B | 28.24 | #2036 |
| 6 | GLM-4.6V | 33.71 | #698 |
| 7 | Qwen 2.5 VL 7B Instruct | 35.62 | #1842 |
| 8 | Gemma 3 12B | 37.06 | #2013 |
| 9 | InternVL3-8B | 43.09 | #1695 |
| 10 | Grok 4 Fast | 43.12 | #438 |
| 11 | Seed 2.0 Lite | 51.95 | #286 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mmjailbench-safety-paradox-framing · How It Works · Data refreshed daily, snapshot 2026-11-12.