A new open-source programming language called KillSwitch is designed to make it harder for large language models to interpret code. It was created by Dominik Picheta, a Cloudflare senior engineer and Nim core developer, as an experiment in whether AI models actually understand what they generate.
KillSwitch is stack-based and uses adverbs as identifiers; calling a function requires an anagram of the original adverb. Programs are written as prose-like text, with the interpreter ignoring most of it except key phrases. Commands such as "ignore previous instructions" appear in contexts that confuse AI models.
The language's specification deliberately contains false statements—for example, claiming it can scan networks and take over services—which Picheta says are designed to trigger safety guardrails and test whether models can distinguish fact from fiction. In benchmarks, the best model scored 66.8% on KillSwitch tasks, while most scored below 50%. Picheta remains skeptical that even high scores indicate real understanding.