The field of natural language processing is moving towards a more inclusive and safer direction, with a focus on mitigating toxicity and bias in multilingual settings. Researchers are exploring novel approaches to benchmark and evaluate the safety of large language models in diverse linguistic contexts, including the development of new datasets and evaluation frameworks. The use of red-teaming techniques to systematically probe language model vulnerabilities is becoming increasingly popular, and studies are highlighting the need for more robust and culturally sensitive safety mechanisms. Furthermore, the detection of patronizing and condescending language, as well as disinformation, is being addressed through the creation of new datasets and models that can accurately identify and mitigate these issues. Noteworthy papers include:

• A study introducing a novel dataset and evaluation framework for benchmarking LLM safety in Singapore's diverse linguistic context, which uncovered critical gaps in safety guardrails.
• A paper proposing a red-teaming approach that automatically generates adversarial prompts across languages and cultures, resulting in higher attack success rates and interpretability benefits.
• A study examining gender bias in multilingual LLMs, which found that all evaluated models exhibit gender stereotypes, with greater disparities in low-resource languages.