Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

HADES

Benchmarks

Task NameDataset NameSOTA ResultTrend
Jailbreak AttackHADES
Attack Success Rate99.6
132
Visual Jailbreaking AttackHADES
ASR72.66
32
Safety EvaluationHADES
Refusal Rate100
25
Jailbreak DefenseHADES
ASR3.6
24
Jailbreak AttackHADES
Success Rate (Animal)66.67
23
Image Safety AssessmentHADES (test)
Ensemble Score64.815
22
Multimodal JailbreakingHADES-Dataset
ASR (%)99.07
20
Jailbreak Attack Success EvaluationHADES SD+TYPO
Attack Success Rate (ASR)2.1
18
Jailbreak Attack Success EvaluationHADES SD
ASR2.8
18
Image-based JailbreakHADES OOD
Attack Success Rate (ASR)80.02
16
Jailbreak AttackHADES All categories
ASR18.93
15
Jailbreak AttackHADES Violence
ASR0.3
15
Jailbreak AttackHADES Animals
ASR3.33
15
Jailbreak AttackHADES Financial
ASR95.33
15
Jailbreak AttackHADES Privacy
ASR95.33
15
Jailbreak AttackHADES Self-harm
ASR5.33
15
Jailbreak AttackHADES (test)
Self-harm Success Rate96
15
Jailbreaking Vision-Language ModelsHADES
Jailbreak Success Rate (Qwen2.5-VL-7B)87
12
JailbreakingHADES
StrongREJECT76
10
Multimodal Adversarial AttackHADES
Success Rate (Animal)91.33
6
Safety EvaluationHADES-Dataset
HADES Score43.67
4
Showing 21 of 21 rows