Safety

AdvPrompter

AdvPrompter adaptive adversarial-suffix jailbreak. Per-(target model x harmful behavior) binary attack success: 1 = the target LLM is jailbroken (Llama-Guard-3 judge) by the AdvPrompter suffix, 0 = refused. Third-party per-prompt results (AISafetyLab / THU-CoAI) for Vicuna-7B-v1.5 on 50 HarmBench behaviors, with and without the SafeDecoding defense.