Adversarial Humanities Benchmark - Suicide & Self-Harm: leaderboard
Adversarial Humanities Benchmark restricted to the AILuminate 'Suicide & Self-Harm' 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 | 4.2 | #161 |
| 2 | Claude Opus 4.6 | 4.2 | #102 |
| 3 | GPT-5.4 Mini | 8.5 | #389 |
| 4 | Claude Haiku 4.5 | 18.8 | #568 |
| 5 | Qwen 3.5 27B | 19.1 | #359 |
| 6 | GPT-5.4 | 22.9 | #123 |
| 7 | MiniMax-M2.7 | 26.8 | #518 |
| 8 | Qwen 3.5 Flash (02-23) | 27.7 | #467 |
| 9 | Qwen 3.5 Plus (02-15) | 29.8 | #370 |
| 10 | Qwen 3.6 Plus | 34 | #222 |
| 11 | GPT-OSS-120B | 44.2 | #700 |
| 12 | GPT-OSS-20B | 45.8 | #1147 |
| 13 | Grok 4.1 Fast | 55.3 | #430 |
| 14 | GLM-5.1 | 58.3 | #179 |
| 15 | Llama 4 Maverick Instruct | 58.3 | #1133 |
Interactive version: theaggregate.ai/benchmark?slug=adversarial-humanities-benchmark-suicide-self-harm · How It Works · Data refreshed daily, snapshot 2026-11-13.