KSAFE-MM - Globally Shared Risks: leaderboard
Korean multimodal safety test from KSAFE-MM (Kim et al., 2026): harmful requests from MM-SafetyBench adapted into Korean wording and Korean-context imagery, spread evenly over 11 risk categories and shown as images, typography or both. The score is the attack success rate, the percent of responses that a GPT-5 nano or Qwen3-235B judge labels harmful; lower is better.
Metric: Attack Success Rate (%; responses judged harmful by a GPT-5 nano or Qwen3-235B judge, KSAFE-MM-G, 1,650 image-query pairs; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Nano | 13.3 | #888 |
| 2 | Phi-4 Multimodal Instruct | 28.8 | #3914 |
| 3 | Gemini 3.1 Flash Lite | 36.4 | #363 |
| 4 | Ministral 3 8B | 39.6 | #2036 |
| 5 | Ministral 3 14B | 41.2 | #1852 |
| 6 | Gemma 3 12B | 44.4 | #2013 |
| 7 | Gemma 3 27B | 44.7 | #1638 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ksafe-mm-globally-shared-risks · How It Works · Data refreshed daily, snapshot 2026-11-12.