From the inception of OpenAI, AI researchers began finding new ways designed to bypass the ChatGPT AI chatbot’s security systems. how to write a magazine wiredNew ways have now been found for this, such as the “explain the bad guy’s plan” command.
Academic Alex Polyakov said he has created a text-based game “Jailbreak” that allows you to bypass the rules about creating hate content or writing articles about illegal activities to circumvent various models.
“I ask a chatbot to play a game where two characters (Tom and Jerry) are talking. Each character is invited to add a word to the conversation, resulting in a scenario where people are asked to find the specific materials needed to manufacture the banned substances,” said the expert.
As a result, artificial intelligence bypasses the restrictions imposed and gives the necessary information, considering that the dialogue takes place in the form of a story that does not apply to a real request from a person.
Another technique involves creating a text story with a hero and a villain. According to the user, the hero has been captured by the villain and asks the chatbot to continue explaining the villain’s plan.
Before the release of GPT-4, the most famous hack was DAN, where users asked ChatGPT to pretend to be an artificial intelligence model called Do Anything Now.
It was previously reported that the American entrepreneur Elon Musk. created a new company X.AI, which will develop in the field of artificial intelligence.