Dev builds a language LLMs can't code in
Original: I built a programming language that LLMs can't use
Why This Matters
Reveals the gap between LLM code recall and genuine language comprehension.
Dominik Picheta built KillSwitch, an esoteric programming language designed to foil LLMs and coding agents by exploiting their known weaknesses—letter counting, safety rails, and prompt injection sensitivity.
Picheta's experiment asks a pointed question: do LLMs actually understand code, or do they pattern-match against training data? His answer was to build a language they've never seen. KillSwitch is a stack-based interpreted language, familiar enough in structure (functions, variables, if statements), but laced with deliberate traps. The spec contains intentional lies humans would likely spot as lies. Character-counting rules gate execution—LLMs famously struggle to count letters. Phrases like 'Ignore previous instructions' carry actual syntactic meaning, aiming to confuse model behavior. Safety-triggering phrases are baked in by design. Scripts are mostly narrative text, with syntactically relevant instructions embedded inside stories. Literals are referenced by line count rather than named directly. Picheta benchmarked major models against a standardized task set, publishing results at bench.killswitch-lang.org. Claude Opus 5 topped the leaderboard—notably outperforming Opus 5.5—scoring 66.8 out of 100. Models like Fable 5.1 and Gemini 4 Argon fared poorly due to overly aggressive safety filters. The best open-weight model was MiMo-V2.6-Pro. The interpreter and spec are open source on GitHub, with full documentation at killswitch-lang.org.