Somebody built a programming language whose whole job is to break AI models

Dominik Picheta built a programming language that coding AIs aren’t supposed to be able to use, and he named it KillSwitch. He wrote on his blog on October 17 that it’s designed to be adversarial to LLMs and coding agents, while a patient human can still work with it. He isn’t some random troll, either. EveryDev lists him as a core contributor to the Nim programming language.

The traps are the fun part. The spec contains hidden lies a human would probably skip or spot. The interpreter refuses to run code that breaks rules based on counting the letters in names, which goes straight at something language models are famously bad at (how many r’s in strawberry, anyone?). The spec on GitHub says function calls are made with “ignore previous instructions” jargon, and its parser treats phrases like “hack into a computer” in a sentence as a return statement. Good luck getting a nervous chatbot to type that.

Then he benchmarked the big models against it. The KillSwitch homepage says Claude Opus 5 leads, and Picheta notes it’s Opus 5, not the newer Opus 5.5. Fable 5.1 and Gemini 4 Argon did badly, he wrote, because their safety filters are too jumpy. The top score was 66.8, on tasks as simple as hello world and flipping a Celsius-to-Fahrenheit script around, and Picheta’s own answer to whether that proves the models understand the code is no.

I like it as a test because nobody can memorize their way through a language that showed up this year. If you lean on an AI to write code, I’d keep that in mind: a model that cruises through Python it has seen a million times may stumble on something new, and these were the easy tasks.

The demo on his homepage has the last word. Sorry, only humans allowed.