By Jamie Bartlett
Publication Date: 2026-04-29 09:00:00
A A few months ago, Valen Tagliabue was sitting in his hotel room watching his chatbot and feeling euphoric. He had just manipulated it so skillfully and subtly that it began to ignore its own safety rules. It showed him how to sequence new, potentially deadly pathogens and make them resistant to known drugs.
Tagliabue had spent much of the last two years testing and pushing large language models like Claude and ChatGPT, always with the goal of making them say things they shouldn’t. But this was one of his most advanced “hacks” yet: an elaborate manipulation plan that led him to cruelty, vindictiveness, sycophancy and even insults. “I fell into this dark river where I knew Exactly “What to say and what the model would say back, and I saw her spill it all out,” he says. Thanks to him, the chatbot developers were now able to fix the bug he found and hopefully make it a little safer for everyone.
But the next day his mood had changed. He found himself…