3 citations · 3 across the 4 of their papers we have counts for
1 paper · 1 filter
Asen Dotsinski, Panagiotis Eustratiadis
Prefill attacks are an effective and low-cost jailbreaking method, as they directly insert an acceptance sequence (e.g., "Sure, here is...") at the start of an LLM's output and lea…