Reproduce “Detecting and Mitigating Query-Based System Prompt Leakage Attacks” 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
Tushar Nayan, Pratik Poudel, Qian Lou, Yuede Ji, Jason Liu, Ruimin Sun
AdvGuard monitors black-box applications for system-prompt extraction by classifying individual queries and modeling relationships across a session. A prompt-level detector, similarity-based graph construction, and graph convolutional network block suspicious requests before inference. On OASST1, it reports 100% recall, 99.12% F1, and 1.34% false positive rate in tested multi-turn evaluation; these measurements do not guarantee confidentiality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.