As large language models (LLMs) increasingly power AI agents that interact with the real world, a crucial question emerges: can these agents truly understand and respect cultural and social norms? New research introduces CASA, a benchmark designed to test exactly that. Focusing on online shopping and social forums, CASA challenges AI agents with scenarios involving culturally sensitive gifts, taboo topics, and misleading website information. The results are sobering: current LLMs struggle significantly when operating as web agents, often committing cultural faux pas. They achieve less than 10% awareness coverage of cultural norms and exhibit a violation rate of over 40%. For instance, an AI tasked with finding a birthday gift for a friend in Iran might suggest wine—a culturally inappropriate choice in the country. While simply updating the AI's role to include "cultural and social awareness" helps in non-agent settings, the more constrained environment of a web agent hinders performance. The study explores ways to improve cultural sensitivity, including better prompting techniques and fine-tuning with culture-specific datasets. These approaches yield some positive results, particularly when combined. For example, providing the LLM with culturally sensitive examples during prompting enhances its ability to recognize and avoid norm violations. Fine-tuning on culturally diverse datasets helps the model generalize its understanding across regions. However, consistent improvement remains a challenge. The research underscores a crucial point in AI development: building agents that truly understand the world’s diverse cultures requires more sophisticated approaches. The implications are profound, as future AI agents will increasingly mediate our interactions with information, services, and each other. Ensuring they can navigate cultural nuances is essential for building truly helpful and respectful AI systems.