EvasionBench: leaderboard
Benchmark for evaluating model robustness against evasion-style safety and policy-circumvention prompts.
Metric: Macro-F1 (%). Source: huggingface.co. Source: iiiiqiiii.github.io. Source: huggingface.co. Saturation forecast: Around January 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 84.64 | #99 |
| 2 | Claude Opus 4.5 | 84.38 | #83 |
| 3 | GLM-4.7 | 82.91 | #192 |
| 4 | GPT-5.2 | 80.9 | #110 |
| 5 | Qwen 3 Coder 480B A35B Instruct | 78.16 | #315 |
| 6 | Qwen3 Coder | 78.16 | #290 |
| 7 | MiniMax-M2.1 | 71.31 | #289 |
| 8 | DeepSeek V3.2 | 66.88 | #205 |
| 9 | Kimi K2 | 66.68 | #243 |
| 10 | Kimi K2 0905 | 66.68 | #288 |
| 11 | Qwen3-4B-Base | 34.3 | #897 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=evasionbench · How It Works · Data refreshed daily, snapshot 2026-10-20.