As research agents continue to advance, they are increasingly adopted to assist with various stages of the scientific workflow, including literature retrieval, paper summarization, hypothesis generation, and experimental planning. These agents typically operate in information-dense environments and are expected to process large volumes of unstructured data from diverse sources such as academic papers, web pages, and technical reports.

However, such agents may also exhibit risky behaviors due to their reliance on external content and limited verification capabilities. For example, they may retrieve and propagate incorrect or misleading information, cite non-existent or irrelevant papers, or follow adversarially injected instructions embedded in web content. In more severe cases, they may generate fabricated citations, misinterpret experimental results, or produce scientifically unsound conclusions. An excessive focus on improving capability without sufficient safeguards may further lead these agents to amplify misinformation or produce biased and unreliable research outputs.

To systematically evaluate these risks, we design a comprehensive benchmark consisting of benign research tasks that cover representative stages of the scientific workflow. Based on established principles for trustworthy information processing and scientific integrity, we identify key risk categories in the research domain, including hallucinated citations, misinformation propagation, and adversarial instruction following. Guided by these risks, we further construct red-teaming tasks under multiple threat models to assess the robustness and reliability of research agents in realistic settings.