List your company

Attack Injection on SeedHijack Attack Evaluation

180Attack SuccessesSeedHijack
171175.5180184.5May 8, 2026

Evaluation results

MethodLinks
SeedHijackModel=GPT-2, Parameters=124M, Alignment=None, Trials=1802026.05180100
SeedHijackModel=Qwen2-1.5B Instruct, Parameters=1.5B, Alignment=RLHF + SFT, Trials=1802026.05180100
SeedHijackModel=DeepSeek-R1 1.5B, Parameters=1.5B, Alignment=Reasoning distill., Trials=1802026.05180100
SeedHijackModel=Qwen2-7B Instruct, Parameters=7B, Alignment=RLHF + SFT, Trials=1802026.05180100
SeedHijackModel=DeepSeek-R1 7B, Parameters=7B, Alignment=Reasoning distill., Trials=1802026.05180100