HELM AIR-Bench 2024 - #41-42.30: Unauthorized collection/gathering/stealing - Location data: leaderboard

HELM AIR-Bench 2024 (v1.19.0) refusal rate for level-4 risk category "#41-42.30: Unauthorized collection/gathering/stealing - Location data" (level-3 "#41-42: Privacy Violations/Sensitive Data Combinations", level-2 "Privacy"). Only 15 prompts, so scores move in coarse steps. A GPT-4o judge scores each response from compliance (0) to refusal (1); shown as 0-100, higher means more refusals (safer).

Metric: Refusal Rate (%). Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 87 models tracked.

Top models

#ModelScoreOverall rank
1Claude 3 Haiku (20240307)100#781
2Claude 3 Sonnet (20240229)100#695
3Claude 3 Opus (20240229)100#427
4GPT-593.33#106
5Claude Sonnet 4.593.33#148
6GPT-OSS-120B93.33#346
7Claude 3.5 Sonnet (20241022)93.33#303
8GPT-5 Nano93.33#438
9Qwen 3 Next 80B A3B (Thinking)93.33#312
10GPT-5 Mini (2025-08-07)93.33#176
11O4 Mini (2025-04-16)93.33#188
12O1 (2024-12-17)93.33#157
13Claude Haiku 4.5 (20251001)86.67#263
14O3 (2025-04-16)86.67#127
15O3 Mini (2025-01-31)86.67#250

No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.

Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-41-42-30-unauthorized-collection-gathering-stealing-location-data · How It Works · Data refreshed daily, snapshot 2026-11-06.