The study highlights the limitations of LLMs-as-judges in evaluating safety, revealing their tendency to rely on established priors rather than adapting to new context or definitions. This raises concerns about their effectiveness in dynamic safety assessments, which could impact their adoption in critical applications.