Research Evidence for Dual-Use Risk Evaluation of Large Language Models
Description
This dataset contains research evidence supporting the paper “Beyond the Prompt: Probing Dual-Use Risks in Large Language Models and a Risk-Based Accountability Framework.” It contains qualitative prompt-response records and researcher observations involving GPT, Claude, and Gemini, primarily covering legal and cybersecurity scenarios, together with Gemini API test records from the Oracle risk-sensitive screening prototype. The records preserve prompt conditions, model information where available, observed responses, classifications, and research observations. The dataset is provided to improve transparency and traceability of the observations reported in the associated paper. It is exploratory qualitative evidence and is not intended to represent a statistically representative benchmark of overall model safety or comparative model performance.