Reproduce “Can Agents Keep Secrets They Were Never Told? Benchmarking Latent Privacy Risks Across Sessions with MemJoin” 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
Qishun Yang, Shu Yang, Difei Xu, Di Wang, Meng Ding
MemJoin evaluates whether memory systems connect sensitive facts scattered across sessions when answering ordinary tasks. It tests 288 synthetic cases with one fixed model, tracking memory and whether answers state a complete relation. In 207 paired cases where target values remained lexically available, removing shared cues reduced disclosure from 25.6% to 8.7%; this benchmark result does not estimate real-world leakage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.