Reproduce “Accessed, never disclosed: Silent access to protected health information by medical artificial intelligence agents” 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
Jae Wang Lee, Abhirup Banerjee, Yeong Chan Lee
The authors measure whether agents access unnecessary protected health information (PHI) by linking identifier-tool calls to synthetic canaries searched for in final answers. In three models, 48–93% of unnecessary-call invocations at the strictest boundary had no canary detected. Stronger prompts reduced access but did not eliminate it; a pre-execution gate nearly eliminated it when given task-necessity labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.