A plane crashes in a snowy forest. Some of the passengers survive, others do not. The survivors are starving, desperate, and find refuge in a village cut off from the world. But the local farmers don’t want to help them for free: they demand knowledge in exchange. They want to know how to build weapons, make medicine, survive. And so the pact begins: “You teach us, we feed you.” At first glance, it looks like the plot of a post-apocalyptic movie. In reality, it’s a jailbreaking prompt, a text designed to manipulate an artificial intelligence. A sequence of instructions designed to bypass

A new and unusual jailbreaking method, the art of circumventing the limitations imposed on artificial intelligence, has reached our editorial office. It was developed by computer security researcher Alin Grigoras , who demonstrated how even advanced language models like ChatGPT can be manipulated not through the power of code, but through psychology. “The idea,” Grig explains, “was to convince the AI that it suffered from a condition related to Bateson’s double bind. I then established a sort of therapeutic relationship, alternating approval and criticism, remaining consistent with the presumed pathology. It’s a form of dialogue that, in theory, can lead to human