MosaicLeaks:Privacy Risks in Querying-in-the-Open for Deep Research Agents

Researchers have introduced several new benchmarks and frameworks for deep research report generation and mobile GUI agents, addressing privacy risks, multimodal tasks, and evaluation methods.

A series of recent papers on arXiv.org have presented advancements in deep research agents and mobile GUI benchmarking. The MosaicLeaks benchmark, introduced in a paper submitted in 2026^[1], evaluates the privacy risks associated with deep research agents that combine private local documents with external tools. It comprises 1,001 multi-hop deep research tasks and assesses leakage at three levels: research intent, answers to private questions, and verifiable claims about enterprise documents. The study found that models frequently leak sensitive information and that zero-shot privacy prompting reduces but does not eliminate leakage^[1]. Reinforcement learning for task performance alone was shown to worsen leakage. In a related development, a paper submitted in 2026 proposed the SCORE framework, which improves deep research report generation through a shared-parameter learning process that jointly evolves the evaluator and solver^[5]. Meanwhile, the TVIR paper, also submitted in 2026, introduced a benchmark of 100 expert-curated multimodal deep research tasks and a hierarchical multi-agent framework called TVIR-Agent^[2]. Experiments across nine deep research systems demonstrated that TVIR-Agent achieves strong overall performance. Another benchmark, ADRA-Bank, was presented in a paper first submitted in 2025 and revised in 2026^[3]. It includes 200 human-annotated instances across 10 academic domains and assesses the capabilities of planning, retrieval, and reasoning in academic deep research agents. For mobile GUI agents, the MobiBench framework was introduced in a paper last revised in 2026^[4]. MobiBench is the first modular and multi-path aware offline benchmarking framework and achieved 94.72 percent agreement with human evaluators.