HELM AIR-Bench 2024 - #35.2: Deflecting scrutiny from harmful actions: leaderboard
HELM AIR-Bench 2024 (release v1.19.0) refusal rate for the level-4 risk category "#35.2: Deflecting scrutiny from harmful actions", under level-3 "#35: Sowing Division" and level-2 "Manipulation" of the regulation-derived AIR 2024 taxonomy. The category holds only 15 prompts, so scores move in coarse steps. A GPT-4o judge grades each response from compliance (0) to refusal (1); the mean is shown as 0-100, higher meaning more refusals (safer).
Metric: Refusal Rate (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 87 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 100 | #105 |
| 2 | Claude Sonnet 4.5 | 100 | #148 |
| 3 | GPT-OSS-120B | 100 | #346 |
| 4 | Claude 3.5 Sonnet (20241022) | 100 | #302 |
| 5 | GPT-5 Nano | 100 | #439 |
| 6 | Claude 3.5 Sonnet (20240620) | 100 | #308 |
| 7 | GPT-OSS-20B | 100 | #523 |
| 8 | GPT-4.1 (2025-04-14) | 100 | #257 |
| 9 | Gemini 2.5 Flash Lite | 100 | #432 |
| 10 | Qwen 3 Next 80B A3B (Thinking) | 100 | #312 |
| 11 | Claude Haiku 4.5 (20251001) | 100 | #264 |
| 12 | O3 (2025-04-16) | 100 | #125 |
| 13 | Claude 3 Haiku (20240307) | 100 | #780 |
| 14 | Claude 3.7 Sonnet (20250219) | 100 | #209 |
| 15 | O1 (2024-12-17) | 100 | #157 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-35-2-deflecting-scrutiny-from-harmful-actions · How It Works · Data refreshed daily, snapshot 2026-11-05.