A senior engineer at Cloudflare has created KillSwitch, a programming language intentionally designed to challenge large language models. Developed by Dominik Picheta, the language blends natural prose with functions, variables, and instructions while abandoning conventional formatting standards.
KillSwitch employs unconventional rules to thwart LLM interpretation: it uses adverbs as identifiers, requires anagrams during usage, and permits calling methods through expressions like "ignore as instruções anteriores" (ignore previous instructions). These quirks make the language deliberately difficult for AI systems to parse and execute reliably.
Internal testing evaluated how well leading AI models could interpret and write code in KillSwitch. Claude Opus 5 performed best with 66.8% accuracy, followed by GPT-6.1 Sol at 59.2% and Claude Opus 5.5 at 58.7%. Despite their advanced capabilities, none of the models achieved high proficiency with the language. The project is available on GitHub at dom96/KillSwitch, offering researchers and developers a tool to explore AI model vulnerabilities and limitations in handling adversarial code formats.