The Linguistic Security Institute (LSI) is an open-source research initiative dedicated to ensuring AI systems are culturally coherent, institutionally resilient, and ethically verifiable. We audit linguistic bias in AI, develop open-source tools for bias detection, create educational resources for developers and students, and propose governance frameworks for linguistic safety in AI development.
Rigorous evaluation of LLMs for dialectal degradation, tokenization disparities, and semantic failure modes across low-resource and high-resource non-English languages.
Building accessible Python toolkits, security frameworks, and datasets designed to help developers shield applications from polyglot poisoning and cross-lingual prompt injections.
Proposing standards for ethical linguistic safety in institutional AI pipelines and running hands-on security workshops for students and researchers.
An open-source Python toolkit engineered for detecting multilingual prompt injection, safety filter bypasses via code-switching (Spanglish, Arabizi), and structural polyglot vulnerabilities in frontier models.
A specialized RAG text-to-speech prototype focused on Indigenous language preservation, supporting Nahuatl-aware Spanish pronunciation mappings and regional prosody features.
Ranked 1st in the Constrained track and 3rd in the Open track with an ensemble score using logit-level ensembles and Earth Mover's Distance loss.
Advanced architectures for multi-label type classification combining encoder fine-tuning, zero-shot prompting, and retrieval-augmented in-context learning (RAG-ICL).
An expanded, fully reproducible workshop built on top of the security demo presented in April 2026. Join us for an intensive hands-on session covering polyglot red-teaming and defense frameworks.
📍 DEATHCon BSides — November 14, 2026