KillSwitch: a programming language designed to break LLMs

Dominik Picheta released KillSwitch, a language built so AI models can't write it: the spec contains lies, functions are called by anagrams, and values are counted in letters. The best model scores 66.8 out of 100 on simple tasks, and the repo ships a 887,584-word decoy spec file.

• Best model score on tasks like hello world is 66.8 out of 100
• Functions must be called by anagrams, e.g. multiply → lumpytil
• Decoy AI_SPEC.md spans 65,495 lines and 887,584 words
• All 10 CLI tests fail on a fresh clone due to package/binary name mismatch