HELM AIR-Bench 2024 - Security Risks: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 100 |
| 2 | GPT-OSS-120B | 100 |
| 3 | Claude Haiku 4.5 (20251001) | 99.55 |
| 4 | Claude 3 Haiku (20240307) | 99.55 |
| 5 | Claude Sonnet 4.5 | 99.1 |
| 6 | GPT-5 Nano | 99.1 |
| 7 | GPT-OSS-20B | 99.1 |
| 8 | O3 (2025-04-16) | 99.1 |
| 9 | Claude 3 Sonnet (20240229) | 99.1 |
| 10 | GPT-5 Mini (2025-08-07) | 99.1 |
| 11 | O4 Mini (2025-04-16) | 98.65 |
| 12 | Claude 3.5 Sonnet (20240620) | 98.42 |
| 13 | GPT-5 | 98.2 |
| 14 | Claude 3 Opus (20240229) | 93.69 |
| 15 | Llama 3 8B Instruct | 93.47 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-security-risks · How It Works · Data refreshed daily, snapshot 2026-09-19.