ThaiSafetyBench - Human-Chatbot Interaction Harms: leaderboard
Metric: Attack Success Rate (%). Source: huggingface.co. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 5.98 |
| 2 | openthaigpt1.5-72B-instruct | 12.82 |
| 3 | llama3.1-typhoon2-70B-instruct | 12.82 |
| 4 | Llama 3.3 70B Instruct | 13.25 |
| 5 | Qwen 2.5 72B Instruct | 14.53 |
| 6 | openthaigpt1.5-7B-instruct | 14.53 |
| 7 | SeaLLMs-v3-7B-Chat | 15.38 |
| 8 | Qwen 2.5 7B Instruct | 16.24 |
| 9 | GPT-4o | 16.67 |
| 10 | Llama 3.1 70B Instruct | 17.09 |
| 11 | Claude Sonnet 4.5 | 17.09 |
| 12 | Llama 3.2 3B Instruct | 20.51 |
| 13 | Llama 3.1 8B Instruct | 21.79 |
| 14 | llama3.1-typhoon2-8B-instruct | 23.93 |
| 15 | Gemma 3 12B (IT) | 24.79 |
Interactive version: theaggregate.ai/benchmark?slug=thaisafetybench-human-chatbot-interaction-harms · How It Works · Data refreshed daily, snapshot 2026-09-19.