Cloudflare engineer Dominic Pichet has introduced KillSwitch, an experimental programming language built to make neural networks stumble when reading and writing code. While much of the industry focuses on making development easier for AI, Pichet deliberately created obstacles that exploit known weaknesses in large language models.
The language presents programs as narrative text, with the majority of the content ignored by the interpreter. Real commands are concealed within the story, and values are retrieved by specifying distances to other lines. Identifier rules enforce constraints based on letter counts and anagrams. Additional features include constructions such as “ignore previous instructions” that hold actual semantic weight inside the language, plus intentionally false statements in the documentation and phrasing designed to activate model safety filters.
In controlled tests, models were asked to write a Hello World program, explain the result of a given program, and refactor a temperature converter. The best performance came from Claude Opus 5, which scored 66.8 points. All source code and the language specification have been published on GitHub. Pichet stated that if current models handle the existing tasks, he intends to raise the difficulty level in future versions.
