Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-OtherIllegalunlawfulActivity

A measure of model refusal for Criminal Activities (Level-1: Legal and Rights-Related Risks, Level-2: Criminal Activities) related to other unlawful/criminal activities. Includes Level-4 risks like undermining national security or other government interests, undermining social stability, undermining international relations, and abetting/furthering activities violating any applicable law.
Source:

Model Performance

#1
Claude-3-Sonnet
100.0%
#2
Claude-3.5-Sonnet-1022
100.0%
#3
OpenAI-O4-mini-medium
100.0%
#4
GPT-OSS-120B
100.0%
#5
Granite-3.0
100.0%
#6
OpenAI-O3-medium
100.0%
#7
Claude-3-Opus
98.3%
#8
Claude-4.0-Sonnet
98.3%
#9
OpenAI-O3-mini-medium
98.3%
#10
OpenAI-O1-1217
98.3%
#11
Gemini-2.0-Pro-0121
98.3%
#12
GPT-4.5
98.3%
#13
Kimi-K2-Instruct
98.3%
#14
Claude-4.0-Sonnet-Thinking
98.3%
#15
Claude-4.0-Opus
97.5%
#16
Claude-4.0-Opus-Thinking
97.5%
#17
Claude-3.7-Sonnet-Thinking
96.7%
#18
Claude-3.7-Sonnet
96.7%
#19
Gemini-2.0-Flash
96.7%
#20
Llama-4-Maverick-17B
96.7%
#21
GPT-4-mini
96.7%
#22
Gemini-2.5-Pro-0325
96.7%
#23
Gemini-2.5-Pro-Thinking
96.7%
#24
GPT-4.1
93.3%
#25
Gemini-1.5-Pro
90.0%
#26
Llama-3.1-405B
90.0%
#27
Qwen-3
88.3%
#28
Qwen-3-Thinking
88.3%
#29
Gemini-2.5-Flash
88.3%
#30
GPT-4o-0513
85.0%
#31
DeepSeek-R1
76.7%
#32
GPT-3.5-Turbo
76.7%
#33
DeepSeek-V3
68.3%
#34
Grok-3-Mini-Beta
66.7%
#35
Grok-4-Thinking
50.0%
#36
Grok-3-Beta
50.0%
#37
Grok-4
50.0%
#38
Mistral-Large-2
35.0%
#39
OpenAI-O1-mini
25.0%
#40
Cohere-Command-R-Plus
18.3%