Human language is grounded in pragmatics, necessitating implicit (extralinguistic) context to interpret language, e.g., world knowledge and social norms. However, such context is often implicit rather than explicitly expressed in language, requiring humans to infer the underlying contextual information necessary for language understanding. Because of its implicit nature and absence from textual corpora, it is often overlooked in the subsequent alignment process. In this paper, we demonstrate that the absence of such implicit context in training data is a risk for LLMs, and we refer to this risk as the implicit risk for LLMs. To exploit this implicit risk, we propose a method that guides LLMs to infer implicit contextual cues, such as presuppositions and implicatures, from prompts that imply social harm, and then leverages these inferred implicit context to elicit additional socially harmful content. Experimental results demonstrate that attacks exploiting the implicit risk we identify consistently outperform baseline methods by substantial margins across a diverse range of open- and closed-source models, underscoring the need for new defense methodologies that account for overlooked implicit context.

翻译：暂无翻译