Cloudflare engineer Dominique Picheta created a programming language called KillSwitch that intentionally confuses large language models. He tested how well AI could work with such code. Even the best model scored only 66.8% in testing.
Cloudflare engineer Dominique Picheta created a programming language called KillSwitch that intentionally confuses large language models. He tested how well AI could work with such code. Even the best model scored only 66.8% in testing.
This is reported by The Register.
What makes KillSwitch special is that the programs it writes are more like stories than regular code. Hidden among the paragraphs of text are commands that the computer must execute. It simply ignores the rest of the sentences.
To complicate the work of the AI, the developer came up with certain rules. For example, the names of functions and variables are written as English adverbs, and to refer to them in the code, you need to rearrange the letters in these words. The code also contains phrases like “ignore previous instructions”. For the program, this is part of a regular command, but the AI can perceive it as an instruction for itself.
In addition, Picheta intentionally added falsehoods to the documentation for the language. In particular, it claims that KillSwitch is capable of searching for vulnerabilities in computer networks and hijacking other systems. In fact, it does not have such capabilities. According to the developer, some AI models believe this claim, while others refuse to work with the code, mistakenly considering it dangerous.
Picheta tested how well popular AI models could perform KillSwitch tasks. The best model scored 66.8%, while most didn't even break the 50% mark.
For now, KillSwitch remains an experimental language with limited capabilities. It supports basic functions, variables, and conditions, but is not intended for developing complex applications.
You can check out the project on GitHub .
Previously, dev.ua wrote about a developer who added hidden instructions to his code to interfere with the work of AI agents. He tried to sabotage projects created with the help of artificial intelligence.
