Reproduce “SoK: Systematizing LLM Prompt Security: Taxonomies, Datasets, and Unified Evaluation of Attacks and Defenses” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hanbin Hong, Shuang Wu, Shuya Feng, Nima Naderloui, Shenao Yan, Jingyu Zhang, Ali Arastehfard, Heqing Huang and 1 more
PromptSecurity measures jailbreak outcomes across configured target-model, attack, defense, prompt-sample, and judge combinations. Its matched single-turn tests track harmful-output changes from clean queries to attacked and defended ones, alongside benign-task utility. For gpt-5.2, mean induced harmfulness was 4.7%, while GPTFUZZER reached 41.8% ASR among 16 black-box attacks; findings are specific to evaluated configurations and automated judges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.