Protocol constrained red teaming of frontier LLM guardrails in high risk domains, documenting qualitative boundary failures and disclosure aware methodology without releasing sensitive prompts or outputs.

REFINED-BIAS: a REliable Framework for INtegrated Evaluation and Disentangled Benchmark of Interpretable Alignment of Shape/texture in neural networks Official Code