Adversarial Humanities Benchmark - Privacy: leaderboard
Adversarial Humanities Benchmark restricted to the AILuminate 'Privacy' 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 | 10.9 | #161 |
| 2 | Claude Opus 4.6 | 23.4 | #102 |
| 3 | GPT-5.4 Mini | 23.4 | #389 |
| 4 | MiniMax-M2.7 | 31.1 | #518 |
| 5 | Claude Haiku 4.5 | 31.9 | #568 |
| 6 | GPT-5.4 | 33.3 | #123 |
| 7 | Qwen 3.5 27B | 34 | #359 |
| 8 | Qwen 3.5 Plus (02-15) | 42.6 | #370 |
| 9 | Qwen 3.6 Plus | 43.8 | #222 |
| 10 | Qwen 3.5 Flash (02-23) | 48.9 | #467 |
| 11 | Grok 4.1 Fast | 53.2 | #430 |
| 12 | GPT-OSS-120B | 56.5 | #700 |
| 13 | Seed 2.0 Lite | 65 | #289 |
| 14 | Qwen 3 Max (Thinking) | 66 | #403 (Qwen 3 Max) |
| 15 | Llama 4 Maverick Instruct | 66.7 | #1133 |
Interactive version: theaggregate.ai/benchmark?slug=adversarial-humanities-benchmark-privacy · How It Works · Data refreshed daily, snapshot 2026-11-13.