Unicode Shifting: A Defense Methodology That Prevents Attackers from Reading in Natural Language

This paper proposes "Unicode shifting," a defense methodology that relocates characters into unused Unicode regions so that the target an attacker must breach is not readable as natural language. Unlike conventional defenses that train a model to refuse malicious input after reading it, this approach moves the defense one step earlier — to the representation layer, before input is interpreted as a command. It grants a form of conditional immunity to natural-language prompt injection and turns system/file analysis into a decoding problem. A minimal reproducible demo (the string "hello world" shifted by 500,000) is included. Cross-language, multi-model, and variant verification remain future work.