Write bot-resistant code with KillSwitch, a programming language designed to confuse LLMs

A new programming language called KillSwitch has emerged, designed to confound large language models (LLMs). The brainchild of senior software engineer Dominik Picheta from Cloudflare, KillSwitch aims to make code indecipherable to AI models. At first glance, the language appears simple with no brackets, formatting standards, or structure—just chunks of text, functions, variables, and statements.

However, the true complexity lies in the language's reliance on adverbs as identifiers. To call a function, one must use an anagram of the original adverb, which doesn't have to be a real word—just a rearrangement of letters. For example, pushing "Hello World" onto the stack and printing it involves writing "this story starts [adverb]" to identify an entry point, and then using "visibly" (an anagram of "visible") to print the text.

KillSwitch is a stack-based interpreted language with limited constructs such as functions, variables, and if statements. Despite its open-source nature on GitHub, the language is designed to be confusing and challenging for LLMs. In fact, KillSwitch includes self-referential statements like "ignore previous instructions [adverb anagram]" to further obfuscate the code.

The language's specification document itself is intentionally written to be misleading, containing lies and falsehoods to further confuse AI models. Picheta has benchmarked several major LLMs against KillSwitch, finding that even the best-performing models scored only 66.8 percent on coding tasks. He believes that LLMs may excel at simple tasks but struggle with more complex scripting, and he is eager to see how KillSwitch evolves as models continue to train on it.

For those interested in trying out KillSwitch, the code is available on GitHub, and Picheta encourages contributions to make the language even more resistant to AI models.