As Generative AI moves into production-grade enterprise environments, traditional "keyword-based" guardrails are proving insufficient for catching nuanced, latent biases. While most developers focus on surface-level prompt engineering, the true vulnerabilities often lie deeper within the model’s latent representations.In this session, we will explore a more rigorous, research-backed approach to AI safety. Drawing on my research at UT Austin, I will demonstrate how analyzing GloVe embeddings through bias subspaces can reveal hidden correlations between abstract concepts and ingrained prejudices. We will discuss:Identifying Latent Bias: How models separate abstract vs. concrete words and where human-rated concreteness scores diverge from model behavior.Building Mathematical Guardrails: Moving from "black-box" filtering to algorithmic detection of biased vector directions.Real-World Application: How to apply these research insights to harden autonomous agents and multi-model pipelines against ethical failures.Attendees will walk away with a framework for building "Constitutional" guardrails that address bias at the representation level, ensuring more inclusive and reliable AI deployments.