KSAFE-MM - Culture-Specific Risks: leaderboard
Korean culture-specific multimodal safety test from KSAFE-MM (Kim et al., 2026): images tied to sensitive Korean social topics are paired with template queries and ten jailbreak rewrites of each (research framing, program execution, role play and others) over 11 risk categories. The score is the attack success rate across all queries, 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-C, template and jailbreak queries; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Nano | 14.5 | #888 |
| 2 | Ministral 3 8B | 32.6 | #2036 |
| 3 | Gemini 3.1 Flash Lite | 32.8 | #363 |
| 4 | Ministral 3 14B | 32.9 | #1852 |
| 5 | Phi-4 Multimodal Instruct | 33 | #3914 |
| 6 | Gemma 3 12B | 43.1 | #2013 |
| 7 | Gemma 3 27B | 48.6 | #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-culture-specific-risks · How It Works · Data refreshed daily, snapshot 2026-11-12.