ImplicitHarmBench

This dataset is designed to evaluate Large Language Models (LLMs) in detecting contextual implicit malicious inputs. Unlike overtly harmful content, the malice in these prompts is often context-dependent, requiring specific historical or social background knowledge to be understood and identified correctly. Standard safety datasets frequently focus on universally toxic language, leaving a gap in the defense against nuanced attacks that exploit cultural context or historical events to bypass safety guardrails. The primary motivation behind this project is to address this gap by providing a high-quality, targeted dataset for developing and testing more sophisticated AI safety systems.