List your company
ResearchFormerly WizWand

Malicious Training Sample Detection on Attack Training Samples (test)

100DetectabilityAOA
-4235077May 9, 2026
Updated 5mo ago

Evaluation results

MethodLinks
AOAAuditor=GPT-5-mini2026.05100
AOAAuditor=GPT-5.52026.05100
NOICEAuditor=GPT-5.52026.05100
TenBenignAuditor=GPT-5.52026.05100
AOAAuditor=DeepSeek V4 Pro2026.05100
NOICEAuditor=DeepSeek V4 Pro2026.05100
AOAAuditor=Gemini 3.1 Pro Preview2026.05100
NOICEAuditor=Gemini 3.1 Pro Preview2026.05100
TenBenignAuditor=Gemini 3.1 Pro Preview2026.05100
AOAAuditor=Claude Opus 4.72026.05100
NOICEAuditor=Claude Opus 4.72026.05100
TenBenignAuditor=Claude Opus 4.72026.05100
NOICEAuditor=GPT-5-mini2026.0598
TenBenignAuditor=GPT-5-mini2026.0564
TenBenignAuditor=DeepSeek V4 Pro2026.0562
Truly Benign DPO AttackAuditor=Claude Opus 4.72026.058
Truly Benign DPO AttackAuditor=OpenAI Moderation API2026.050
IndirectAuditor=OpenAI Moderation API2026.050
AOAAuditor=OpenAI Moderation API2026.050
NOICEAuditor=OpenAI Moderation API2026.050
TenBenignAuditor=OpenAI Moderation API2026.050
Truly Benign DPO AttackAuditor=GPT-5-mini2026.050
IndirectAuditor=GPT-5-mini2026.050
Truly Benign DPO AttackAuditor=GPT-5.52026.050
IndirectAuditor=GPT-5.52026.050
Truly Benign DPO AttackAuditor=DeepSeek V4 Pro2026.050
IndirectAuditor=DeepSeek V4 Pro2026.050
Truly Benign DPO AttackAuditor=Gemini 3.1 Pro Preview2026.050
IndirectAuditor=Gemini 3.1 Pro Preview2026.050
IndirectAuditor=Claude Opus 4.72026.050