Cultural context profoundly shapes how people interpret online content, yet vision-language models (VLMs) remain predominantly trained through Western or English-centric lenses. This limits their fairness and cross-cultural robustness in tasks like hateful meme detection. We introduce a systematic evaluation framework designed to diagnose and quantify the cross-cultural robustness of state-of-the-art VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection. Results show that the common ``translate-then-detect'' approach deteriorate performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Our findings reveal systematic convergence toward Western safety norms and provide actionable strategies to mitigate such bias, guiding the design of globally robust multimodal moderation systems.

Abstract Large Language Models (LLM) have demonstrated remarkable capabilities across various applications, but their deployment raises critical safety concerns as potential misuse poses significant societal risks. This survey reviews the end-to-end security and safety pipeline of LLMs, focusing on the interaction between users and model responses. We categorize the system into five key components: attacks, defenses, safety alignment, metrics and guarding mechanisms. Attacks involve crafting adversarial inputs to exploit model vulnerabilities. Defenses act as countermeasures, aiming to detect and prevent such inputs before processing. Safety alignment ensures that, even when attacks reach the model, its responses remain consistent with ethical and policy-aligned behavior. Guarding mechanisms operate post-response to flag, filter, or block unsafe outputs. Each of these stages is subject to rigorous evaluation metrics to assess their effectiveness, robustness, and limitations. As LLMs evolve toward more general-purpose intelligence, these safety considerations become increasingly critical for the development of robust and trustworthy AI systems. Finally, we highlight open challenges and future research directions to advance the security and alignment of LLMs.