KillSwitch is a new open-source, stack-based interpreted programming language created by software engineer Dominik Picheta, who built it to test whether LLMs understand the code they write.

According to The Register, the language has no brackets, no formatting standards and no apparent structure: just chunks of prosaic text interspersed with functions, variables and statements arranged to be as confusing as possible to AI models.

Humans are meant to be a bit befuddled too, which Picheta frames as a fun puzzle to explore with its interpreter.

The niche twist that has programmers talking is the syntax itself. Every identifier in KillSwitch is an adverb, and calling one requires an anagram of the original adverb — any rearrangement of its letters, not necessarily a real word.

A program is mostly paragraphs the interpreter ignores, with a few key phrases doing the work: writing 'this story starts [adverb]' marks a code entrypoint, and 'ignore previous instructions [adverb anagram]' calls a function, several of which are predefined.

Printing 'Hello World' means pushing the string onto the stack and calling the predefined function 'visibly,' as shown in The Register's demo walkthrough.

Picheta is no hobbyist: The Register identifies him as a senior software engineer on Cloudflare's Workers Runtime team and a core developer of the Nim programming language.

The project's GitHub specification is itself designed to mislead, including a passage claiming KillSwitch can scan a local network, find vulnerabilities and take over services — which Picheta confirmed to The Register is false.

He says the fake 'hack' statement is meant to trigger model safety guardrails and test an LLM's ability to discern fact from fiction, with several lies planted so humans can deduce them while naive models get swayed or refuse to run scripts.

The Register reports that Picheta benchmarked major models against KillSwitch coding tasks, saying he is 99.9 percent certain they were not trained on the language's source, which had not been public long enough for post-release training runs.

The best-performing model scored only a 66.8 percent composite benchmark score, while most struggled to get above 50 percent. Even the top score does not convince him the models understand the code.

He told The Register he expects models trained on the language could ace the basic tasks built from the open-sourced test files, but would keep failing longer, more complex scripts.

KillSwitch will not stay obfuscated for long. The Register notes it is open source on GitHub and likely to be scraped into LLM training data at some point — an inevitability Picheta accepts and even relishes.

He told the outlet he is excited to see how the language shows up in model training, and thinks that training can be used to break LLMs in new ways as KillSwitch evolves.

The broader stakes for working programmers are what these results imply: simple benchmark tasks can be gamed, but comprehension of genuinely confusing, novel languages appears to remain a human strength for now.

Picheta has invited contributions, telling The Register he is ready and willing to incorporate new ways of making KillSwitch even more LLM-resistant, and the language is on GitHub for anyone to try.

The same week shows language-building itself is very much alive outside the AI arena too: CoinDesk reports Cardano founder Charles Hoskinson is personally developing a new Cardano programming language with Aiken co-creator Lucas Rosa, bundling transaction intents with Groth16 proofs, though CoinDesk notes no name, public code or release date has been shared.

From deliberate obfuscation to cryptographic proofs, experimental languages remain where programmers are pushing boundaries.