The Implications of Linguistic Illegibility for LLM Security

Large language models generate natural language, but their internal computations are mathematical operations over activation spaces, creating a gap between linguistic outputs and actual model reasoning. This “linguistic illegibility” means that security checks relying on a model’s self-reported language, such as chain-of-thought monitoring or constitutional self-critique, cannot be fully reliable. Taint-tracking policies and robust sandboxing techniques that do not depend on linguistic state are proposed as essential safeguards.