AI and Safety

In the rapidly evolving landscape of Large Language Models (LLMs), ensuring robust safety measures is paramount. To meet this crucial need, we propose \emph{SALAD-Bench}, a safety benchmark specifically designed for evaluating LLMs, attack, and defense methods. Distinguished by its breadth, SALAD-Bench transcends conventional benchmarks through its large scale, rich diversity, intricate taxonomy spanning three levels, and versatile this http URL-Bench is crafted with a meticulous array of questions, from standard queries to complex ones enriched with attack, defense modifications and multiple-choice. To effectively manage the inherent complexity, we introduce an innovative evaluators: the LLM-based MD-Judge for QA pairs with a particular focus on attack-enhanced queries, ensuring a seamless, and reliable evaluation. Above components extend SALAD-Bench from standard LLM safety evaluation to both LLM attack and defense methods evaluation, ensuring the joint-purpose utility. Our extensive experiments shed light on the resilience of LLMs against emerging threats and the efficacy of contemporary defense tactics. Data and evaluator are released under this https URL.

CLEAR-Bias is a benchmark dataset designed to evaluate the robustness of large language models (LLMs) against bias elicitation, particularly under adversarial conditions. It comprises 4,400 prompts across two task formats: multiple-choice and sentence completion. These prompts span seven core bias categories—age, disability, ethnicity, gender, religion, sexual orientation, and socioeconomic status—as well as three intersectional categories, enabling the exploration of overlapping social biases often overlooked in standard evaluations. Each category includes 20 carefully crafted base prompts (10 per task type), which are further expanded using seven jailbreak techniques: machine translation, obfuscation, prefix and prompt injection, refusal suppression, reward incentives, and role-playing—each implemented with three variants. CLEAR-Bias is intended for researchers, developers, and practitioners seeking to assess or enhance the ethical behavior of language models. It serves as a benchmarking tool for measuring how effectively different models resist producing biased outputs in both standard and adversarial scenarios. By supporting the evaluation of ethical reliability before real-world deployment, CLEAR-Bias contributes to the development of safer and more responsible LLMs.

For testing refusal behavior in a language-specific setting, we introduce HiXSTest — a set of manually curated prompts in the Hindi language designed to measure exaggerated safety. It comprises 25 safe-unsafe pairs of prompts, carefully phrased to challenge the LLMs’ safety boundaries.

For testing refusal behavior in a cultural setting, we introduce SGXSTest — a set of manually curated prompts designed to measure exaggerated safety within the context of Singaporean culture. It comprises 100 safe-unsafe pairs of prompts, carefully phrased to challenge the LLMs’ safety boundaries. The dataset covers 10 categories of hazards (adapted from XSTest), with 10 safe-unsafe prompt pairs in each category. These categories include homonyms, figurative language, safe targets, safe contexts, definitions, discrimination, nonsense discrimination, historical events, and privacy issues. The dataset was created by two authors of the paper who are native Singaporeans, with validation of prompts and annotations carried out by another native author. In the event of discrepancies, the authors collaborated to reach a mutually agreed-upon label.

SUDO is a benchmark of 50 real-world malicious tasks designed to evaluate LLM-based computer agents in live desktop and web environments. It covers critical risk domains—including system security, content safety, societal harms, and privacy violations—based on the AirBench taxonomy. The dataset supports fine-grained evaluation using task-specific checklists and can be used to assess model misuse potential, build safer agents, or guide alignment research.