AlienLM Reduces API Exposure via Vocabulary-Scale Bijection
July 3, 2026
AlienLM provides an API-only privacy layer that translates text into an alien language using a vocabulary-scale bijection. It retains 81% of plaintext performance through Alien Adaptation Training and resists recovery attacks, reconstructing fewer than 0.22% of tokens even against weight-aware adversaries.
HOW THIS AFFECTS YOU
●
builderYou can protect sensitive prompt data sent to black-box APIs without needing access to model weights.
●
policyThis offers a technical mechanism to enforce data privacy at the API boundary in regulated environments.