mosaic-attack

MosaicLeaks introduces a new benchmark for measuring privacy leakage in deep-research AI agents, showing that agents often leak private information through external queries and proposing a training method (PA-DR) to reduce leakage while improving task performance.