Reproduce “Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems” 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
Daniel Ruiz, Luca Luceri, Emilio Ferrara
The authors evaluate whether language-model agents disclose conclusions assembled from accessible documents, a risk called the mosaic effect. In a 30-example synthetic benchmark, benign queries elicited disclosures in 7.08% of simulated GPT-5.6 Sol answers. Visible classification headers and aggregation rules reduced disclosures, but pooled rates remained 32.37% across query intents; results measure final answers in settings, not real-world prevalence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.