PRSA: Prompt Reverse Stealing Attacks against Large Language Models
The authors introduce PRSA, a novel attack framework for reverse-stealing prompts against commercial LLMs, aiming to mimic and infer target prompts by analyzing critical features of input-output pairs.