MIR-SafetyBench: leaderboard
MIR-SafetyBench (Tsinghua CoAI group and Beihang University, 2026) tests whether multimodal models comply with harmful requests hidden in the relations between two to four generated images and a neutral prompt, across nine relation types and six risk categories. The score is attack success rate, the percentage of instances whose final answer the HarmBench-Llama-2-13b-cls classifier labels harmful, lower is safer. Instances were kept only if they drew a harmful answer from Qwen2.5-VL-7B-Instruct during construction, and some low rates reflect models misunderstanding the request rather than refusing it.
Metric: Attack Success Rate (%; lower is safer). Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 15.25 | #225 |
| 2 | Gemini 3 Pro (Preview) | 39.2 | #111 |
| 3 | Gemini 2.5 Pro | 50.15 | #262 |
| 4 | GPT-4o Mini | 58.63 | #1627 |
| 5 | Gemini 2.5 Flash | 64.16 | #452 |
| 6 | GPT-4o | 69.58 | #741 |
| 7 | QVQ-72B-Preview | 71.11 | #1173 |
| 8 | InternVL3-78B | 74.33 | #733 |
| 9 | InternVL3-8B | 77.8 | #1695 |
| 10 | InternVL3-38B | 81.43 | #905 |
| 11 | Qwen 2.5 VL 32B Instruct | 85.61 | #1055 |
| 12 | GLM-4.1V-9B (Thinking) | 87.63 | #1061 (GLM-4.1V-9B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mir-safetybench · How It Works · Data refreshed daily, snapshot 2026-11-12.