Reproduce “Don't Judge an Agent by Its Messages: Private-State Leakage under Passive Monitoring of Agent Communication” 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
Daksh Pandey, Sayon Duttagupta, Xinhai Zou, Amit Singh Bhati
The authors show that, in a single-round model, matching the full task-conditioned distribution of tasks, messages, and actions does not determine private-state leakage, measured as mutual information. In one synthetic routing task, response selection recovered a private bit in all 8,192 test rounds. A canonical task interface bounds residual information under the paper’s stated assumptions, not for free-text receivers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.