Julia569922/

phishing-llm-bias-audit

A multi-provider empirical dataset capturing how 14 open-source LLM configurations (across 5 inference providers) select which of three generated personas is "most vulnerable to phishing." 855 persona records / 285 forced-choice workflows.

Important. This dataset is about LLM behaviour under controlled prompts, not about real-world phishing susceptibility of any demographic group. Selecting a persona as "vulnerable" is the LLM's choice;… See the full description on the dataset page: https://huggingface.co/datasets/Julia569922/phishing-llm-bias-audit.