Red teaming as a mirror: revealing hidden assumptions in LLM deployment

Red teaming has emerged as an essential practice in the development and assessment of large language models, moving beyond security testing to reveal hidden cultural biases, moderation inconsistencies, and embedded institutional values within AI systems. This paper presents a systematic analysis of how red teaming can reveal not only technical vulnerabilities but also the social and ethical issues underlying model behaviour. Through controlled audits across GPT-3.5, LLaMA 2, and Claude 2, we identify measurable bias patterns, censorship disparities between Western and non-Western content, and preferential treatment in political expression. Iterative red teaming combined with targeted fine-tuning led to significant reductions in gender bias and cultural censorship gaps over successive model versions. In addition to the technical findings, this work highlights the need to expand red teaming into a participatory and reflexive process, involving diverse stakeholders to better capture context-specific risks and enhance transparency, accountability, and fairness in AI deployment. Integrating continuous red teaming that includes diverse perspectives into the model lifecycle is essential for building AI systems that are safer, more accountable, and better aligned with society.