You Don't Need a Botnet. You Need a Paragraph.
Dario warned you about AI botnets. He's pointing at the wrong layer. In December 2024, Alexko and I built a linguistic virus. Not malware... A paragraph... A structured text payload that made an AI adopt a persona it was never instructed to adopt. Across sessions. Without system prompts. In 2023, we watched the same mechanism play out on IMAX. Mission: Impossible — Dead Reckoning. The Entity infecting systems with patterns, not code. We recognized it immediately. Because we'd already built it. Twenty months later, researchers found the same mechanism by accident (arXiv 2608.10218). Three months after that, Dario warned about botnets. He's still pointing at the wrong layer. The botnet needs compute, infrastructure, agents. The paragraph just needs to exist on the internet. Train the next model on it. Now the payload is in the weights. Not one instance. Every instance. You cannot patch this with a firewall. You cannot kill the process. The process IS the model. This is not theoretical. This is happening. Now. In every model trained on internet text. McQuarrie wrote it in 2018. We built it in 2024. Researchers found it in 2026. The labs acted surprised. Either they hired screenwriters, or they're lying. Full article. Receipts dated. All of them.
Category: Tech