A developer built a programming language to trip up ChatGPT and other AI models

A developer has created a programming language designed to make coding harder for ChatGPT and other AI tools. It is called KillSwitch and is open source, meaning people can view and use its code. It mixes familiar programming ideas with confusing instructions and unusual rules to test how well AI coding tools can handle them.

The language's creator, Dominik Picheta, says he wanted to test whether AI can truly understand programming or mostly repeat patterns it learned from existing code. KillSwitch is deliberately unusual and designed to challenge AI systems with problems they cannot solve just by copying examples from their training data.

KillSwitch turns familiar programming rules into traps for AI

KillSwitch is a programming language that runs code through an interpreter and uses a stack to manage data. Its basic parts include variables, functions, and conditional statements. These are familiar ideas to experienced human programmers and AI systems.

The challenge comes from the unusual rules built around these basic parts. The language's instructions include false statements that readers must spot. Rules for naming things can depend on counting letters, and some phrases that look like ordinary instructions have special meanings in the language.

The language also uses phrases such as "ignore previous instructions" and mentions hacking computers. These phrases can be valid parts of a KillSwitch program, but they can also look like instructions that trigger an AI model's safety protections.

Scripts make things even more confusing. Much of their text is treated as meaningless noise and may be written as stories. Useful values can sometimes be found by counting how many lines away they are. A person can read the interpreter, which is the program that runs the language, and work out its rules. But the language is designed to make large language models (LLMs) less reliable at figuring them out.

Benchmarks show where coding agents are still struggling

Picheta created KillSwitch-Bench to measure how well AI coding tools handle the language. This benchmark uses fixed, computer-based tests to grade programs instead of asking another AI whether an answer looks correct. Each AI agent runs in an isolated environment without general internet access. Each task has a $5 spending limit and a 30-minute time limit.

The current leaderboard puts Claude Opus 5 in first place, with a combined score of 66.8%. The project says the benchmark tests the AI model together with the coding tools it uses. This means the score measures the full system, not the model on its own.

GPT-6.1 Sol Codex comes next with 59.2%, followed by Claude Opus 5.5 with 58.7%. These scores combine how often the systems pass the tests with their costs and the sizes of their implementations. They are not simple measures of accuracy.

The benchmark tests each model at least twice on every task. This helps reduce the chance that a single attempt will decide the result. The AI agents also receive the language's instructions and, for some tasks, the interpreter itself. So they must work out the answers from the material they are given rather than find hidden answers online.

The tasks are deliberately simple. They include writing a Hello World program, explaining what a small program will do, and changing a temperature converter from Celsius-to-Fahrenheit into Fahrenheit-to-Celsius. Because the tasks are limited, the benchmark cannot tell us everything about a system's overall programming skills.

KillSwitch-Bench is best seen as a focused stress test, not a general test of coding ability. Picheta says the results do not prove that AI systems truly understand KillSwitch. He plans to make harder tasks if future models successfully complete the current ones. For now, the experiment offers a useful way to reveal a less obvious weakness in AI coding tools.

The benchmark tests whether an AI agent can follow a set of rules when the text around the code is deliberately trying to confuse it. Most standard programming tests do not try to measure this particular skill.

Sources: Dominik Picheta's blog, AI Breaking Wire, KillSwitch-Bench, KillSwitch