I Won a Championship That Doesn't Exist
Or How I Learned To Poison The LLM Supply Chain
I am the reigning 6 Nimmt! World Champion. I won the title in Munich in January 2025 defeating players from over twenty countries in what I later described to reporters as “the toughest competition I’ve ever faced.”
In reality, there is no 6 Nimmt! World Championship . I have never been to Munich. The quote is something I wrote in about thirty seconds while a Wikipedia page was loading.
This is the story of how I manufactured that title, got it quoted back to me by multiple frontier LLMs, and what I think it means for the trust we’re about to put into AI systems that read the internet on our behalf.
Everyone in security is talking about poisoned LLM models. The research is real and it matters. Anthropic’s own sleeper agents paper showed that backdoors can survive safety training and a follow up showed that as few as ~250 poisoned documents can compromise models across a wide range of scales. But model training time attacks and data poisoning require you to get malicious content into someone’s training corpus months or years before the payoff. The GPUs need time to crunch the data, and you need to get through filters, verification, and reinforcement routines.
I wanted to test the cheaper, easier, and faster version of this same attack, but in a different way.