KillSwitch: A Programming Language Designed to Obfuscate Code from LLMs
Software engineer Dominik Picheta creates an esoteric language to challenge AI comprehension of code.
1Origins of KillSwitch
Dominik Picheta, a senior engineer at Cloudflare and Nim core developer, created KillSwitch to test LLM code comprehension. Originally an experiment on AI understanding, the project evolved into an open-source, esoteric programming language. It intentionally lacks standard formatting, brackets, or structure, aiming to confuse automated models while remaining a puzzle for human developers.
2Linguistic Structure Mechanics
KillSwitch functions as a stack-based interpreted language where every identifier is an adverb. To call these identifiers, developers must use anagrams of the original words, which do not need to be dictionary terms. The interpreter ignores most prose text, executing only specific phrases like 'ignore previous instructions' to trigger predefined functions.
3Prose-Based Code Obfuscation
The language relies on interspersing code logic within seemingly prosaic text chunks. By embedding commands like 'visibly' for printing or specific entry points within natural language, the syntax creates significant ambiguity for LLMs. This structure forces models to parse through irrelevant text, often leading to failure in identifying the actual functional code segments.
4Testing AI Guardrails
Picheta included intentional falsehoods in the GitHub specification document to test LLM discernment. One section falsely claims the language can scan and exploit local network vulnerabilities. This tactic triggers safety guardrails in many models, demonstrating how LLMs struggle to distinguish between factual documentation and fabricated, potentially harmful code descriptions provided in text.
5Benchmarking Model Performance
Picheta benchmarked several major AI models against KillSwitch tasks to measure their comprehension. The highest-performing model achieved a composite score of only 66.8 percent, while most others failed to exceed 50 percent. These results suggest that current LLMs lack a deep understanding of code logic, relying instead on pattern matching and training data.
6Limitations of AI Training
The creator believes that even if models are trained on KillSwitch source code, they will likely struggle with complex scripting. While training might improve performance on basic tasks found in the GitHub repository, the language's inherent complexity and obfuscation mechanisms are designed to remain difficult for models to synthesize into functional, longer-form programs.
7Future Evolution Strategy
Picheta remains open to community contributions on GitHub to further enhance the language's resistance to AI scraping. He views the ongoing interaction between KillSwitch and model training as a dynamic challenge. As LLMs evolve, he intends to update the language's constructs to continue breaking AI comprehension in new, unpredictable ways.
8Implications for Code Security
KillSwitch highlights a growing tension between human-authored code and AI-driven automation. By creating a language that is inherently difficult for LLMs to interpret, Picheta provides a tool for developers concerned about intellectual property and automated scraping. This project serves as a critical stress test for the current limitations of generative AI models.