The Digital Afterlife of Empires: Four Language Models Converge on the Same Imperial Cartography of Writing

This collection of new arXiv papers highlights multiple security-relevant risks from contemporary generative AI and related infrastructure. Key findings: (1) LLMs encode systematic, convergent biases across architectures (language/script, religion, race, gender) driven by historical and corpus-level inequalities, risking misrepresentation and discriminatory outputs; (2) A dedicated medical-safety benchmark (MedHarm) shows aligned and medically-tuned LLMs can still produce unsafe, actionable medical advice across high-risk categories (toxicology, covert poisoning, anesthesia, fetal harm), under