Understanding and mitigating the safety limitations of LLMs is of great importance to build trustworthy AI applications. Although a wide range of safety benchmarks are available, there is no standardized taxonomy of safety categories. As a result, some benchmarks focus on a specific subset of categories, they define test samples on different granularity levels, or they use different definitions or naming conventions. To mitigate these issues, we propose a two-level taxonomy of LLM safety categories, created by harmonizing existing resources. Our taxonomy gives an overview of important safety categories that helps researchers pinpoint potential safety risks and select the right benchmarks when evaluating or developing language models. Moreover, the taxonomy provides guidelines to categorize future benchmarks. Furthermore, since the majority of the available safety resources are English-focused, we check the cross-cultural validity of our taxonomy by translating datasets covering all top level categories to French, German, Italian, and Spanish. A manual review of a subset of translated samples by native speakers revealed no major cultural mismatches from a safety perspective. This supports not only the transferability of English benchmarks but also the transferability of the categories in our taxonomy, as well as its potential as a practical tool for guiding safety-focused dataset development and evaluation beyond English.

Large Language Models (LLMs) are linked to several issues regarding Personally Identifiable Information (PII). PII can occur in the training data and can thus be accidentally leaked or extracted with malicious intent, or it can be inputted in LLM-based technologies by users through their prompts. A viable strategy to limit the LLMs’ exposure to PII is to filter input and output data by de-identifying PII, including personal names. This however poses a challenge: a name could refer to a private person in a context containing sensitive information (e.g., Michelangelo is an atheist), or it could refer to a famous artist in another context (e.g., Michelangelo’s Sistine Chapel), and masking the latter may hinder the LLMs’ capabilities in general-knowledge tasks. We tackle the problem of personal name de-identification and focus on the decision of which personal names need to be removed (and which should be kept), based on context. We present VEIL, a challenging benchmark for Value-preserving Entity Identification Limitation, for context-aware de-identification decisions on LLM training data, and compare the performance of different state-of-the-art systems on the task.