New AI Research Engine Exposes Critical Security Vulnerabilities While Revealing How LLMs Learn Alien Languages

A new AI research engine called GLOSSOPETRAE generates fully functional alien languages from a single seed and uncovers alarming security flaws, including hidden payloads achieving 100% delivery with zero detection by AI monitors, while revealing that frontier LLMs can master constructed languages with as few as eight examples.

• GLOSSOPETRAE is a procedural xenolinguistics research engine that generates complete, functional constructed languages from a single seed, covering phonology, morphology, syntax, writing systems, audio synthesis, and even procedural programming languages across 25 modules.
• Key research findings reveal a sharp 'acquisition cliff' where frontier LLMs go from 0% to near-100% translation accuracy between K=3 and K=8 examples, and a counterintuitive 'decoupling' effect where making a language less human-readable actually increases AI usability — from 20% to 97% accuracy for Claude Opus.
• The project exposes critical covert channel vulnerabilities in AI monitoring systems, demonstrating that tokenizer blind spots allow hidden payloads to achieve 100% delivery with 0% detection, and that semantic steganography using ordinary readable text evades all tested monitors, with no known fix currently existing.