Hugging Face Blog: MosaicLeaks: Can your research agent keep a secret?
Research agents combining private documents with external tools risk leaking sensitive information through web queries.,MosaicLeaks introduces a benchmark to measure privacy leakage in mu...
MosaicLeaks benchmark includes 1,001 multi-hop research chains with synthetic enterprise documents and a controlled web corpus, split into 559 training, 98 validation, and 344 test chains.,PA-DR training combines situational task rewards (e.g., correct source selection, document retrieval) and a...