Source record

Published by Hugging Face Blog (T1.5 source). The original is at https://huggingface.co/blog/ServiceNow/mosaicleaks.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

ServiceNow introduces MosaicLeaks, a benchmark for measuring how deep research agents leak private information through external web queries. Across tested models, agents frequently leaked private data. Their Privacy-Aware Deep Research (PA-DR) RL training method raised strict chain success from 48.7% to 58.7% while cutting leakage from 34.0% to 9.9%.

Frames the mosaic effect as a concrete failure mode for deployed agents and offers a measurable mitigation. Relevant to anyone building or evaluating deep-research systems handling private data.

Cited by

No citations on record.