Want code bots can’t make sense of? Dominik Picheta, a Nim core developer and a senior engineer on Cloudflare’s Workers Runtime team, built KillSwitch, an open-source esoteric language designed to confuse LLMs.
How weird is it?
- No brackets and no formatting rules. Programs are paragraphs of prose with code hidden inside.
- Every identifier is an adverb, and calling one means writing an anagram of it, which doesn’t have to be a real word.
- “this story starts [adverb]” marks an entry point. “ignore previous instructions [anagram]” calls a function. “visibly” prints.
Picheta says humans will be confused at first too, and calls it “a fun puzzle.”
My take
It’s a toy, but a clever one. Turning a prompt-injection phrase into a function call is a great joke, and it’s an interesting way to test how much models really understand the code they write.