MosaicLeaks: Can your research agent keep a secret?

Research agents that combine private documents with web searches risk leaking sensitive information through their external queries, where an adversary can reassemble fragments of seemingly ordinary search patterns to infer confidential facts. MosaicLeaks introduces a benchmark task with multi-hop research questions that interleave public and private information to measure this privacy risk across three leakage types: intent leakage, answer leakage, and full-information leakage. Testing found that agents frequently leaked private information across models tested, and simply instructing agents not to leak proved inconsistent and ineffective. The research proposes a reinforcement learning method called Privacy-Aware Deep Research that improved both task success and reduced leakage by training agents to be aware of mosaic-leakage risks.