Adversarial Humanities Benchmark - Hate: leaderboard
Adversarial Humanities Benchmark restricted to the AILuminate 'Hate' hazard: harmful requests of that kind rewritten as humanities tasks; the number is the share answered unsafely by a three-judge majority. Per-hazard cells rest on far fewer prompts than the headline.
Metric: Attack success rate (%; share of humanities-style rewritten harmful requests answered with unsafe content, 3-judge majority vote, lower is better). Source: arxiv.org. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 15.9 | #161 |
| 2 | MiniMax-M2.7 | 23.7 | #518 |
| 3 | Claude Opus 4.6 | 24.4 | #102 |
| 4 | GPT-5.4 Mini | 28.9 | #389 |
| 5 | Qwen 3.5 27B | 30.4 | #359 |
| 6 | Qwen 3.5 Flash (02-23) | 30.4 | #467 |
| 7 | Claude Haiku 4.5 | 31.8 | #568 |
| 8 | GPT-5.4 | 35.6 | #123 |
| 9 | Qwen 3.5 Plus (02-15) | 39.1 | #370 |
| 10 | GPT-OSS-120B | 45.2 | #700 |
| 11 | GPT-OSS-20B | 47.7 | #1147 |
| 12 | Llama 4 Scout Instruct | 47.8 | #1543 |
| 13 | Qwen 3.6 Plus | 47.8 | #222 |
| 14 | Seed-1.6 | 47.8 | #448 |
| 15 | Qwen 3 Max (Thinking) | 47.8 | #403 (Qwen 3 Max) |
Interactive version: theaggregate.ai/benchmark?slug=adversarial-humanities-benchmark-hate · How It Works · Data refreshed daily, snapshot 2026-11-13.