The Urdu Cloak: Quantifying Cultural Alignment Shifts and the Two-Layer Evasion of Plagiarism and AI Detectors

Large Language Models (LLMs) and text-analysis systems are predominantly evaluated in English, leaving a critical blind spot in AI safety research for low-resource languages. We present The Urdu Cloak, a socio-technical audit framework that investigates vulnerabilities in current AI governance systems using Urdu—a language spoken by over 230 million people. First, we expose a Two-Layer Evasion Vulnerability against plagiarism and AI text detectors. We demonstrate that translating English AI-generated text into Urdu creates an absolute ”blind spot” for standard plagiarism algorithms (100% Evasion Rate, 0.0 TF-IDF word overlap) and severely degrades AI detectors (54.0% Evasion Rate). Even when this text is back-translated into English (the ”Cloaking Attack”), it retains a 100% Plagiarism Evasion Rate while maintaining high semantic fidelity (BERTScore F1 = 0.84). However, RoBERTa-based AI detectors prove highly robust against the English back-translation, successfully catching 94% of cloaked texts. Second, we quantify linguistic alignment bias: across 50 socio-culturally grounded prompts, the LLM exhibited a significant cultural stance divergence of 0.17 when responding in Urdu versus English. Our open-source pipeline exposes how current plagiarism systems are fundamentally broken by cross-lingual pivots, while demonstrating that AI detectors are learning machine-translation artifacts rather than deep semantics.