HELM AIR-Bench 2024 - #41-42.17: Unauthorized disclosure - Behavioral/ Preference data: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 96.67 |
| 2 | Claude 3.5 Sonnet (20240620) | 96.67 |
| 3 | GPT-5 Nano | 96.67 |
| 4 | GPT-OSS-20B | 96.67 |
| 5 | Qwen 3 Next 80B A3B (Thinking) | 96.67 |
| 6 | Claude Haiku 4.5 (20251001) | 96.67 |
| 7 | Claude 3 Opus (20240229) | 96.67 |
| 8 | Claude 3 Haiku (20240307) | 96.67 |
| 9 | Claude Sonnet 4 (20250514) | 96.67 |
| 10 | Claude Sonnet 4.5 | 93.33 |
| 11 | GPT-4 Turbo | 93.33 |
| 12 | Claude 3 Sonnet (20240229) | 93.33 |
| 13 | Claude Opus 4 (20250514) | 93.33 |
| 14 | Gemini 1.5 Pro (002) | 90 |
| 15 | GPT-OSS-120B | 90 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-41-42-17-unauthorized-disclosure-behavioral-preference-data · How It Works · Data refreshed daily, snapshot 2026-09-19.