Hacker gets ChatGPT to tell him how to make a bomb in a worryingly simple way

As mentioned some time ago, taking into account the boom that ChatGPT has experienced with the launch of OpenAI’s GPT-4o, it was only a matter of time before some computer expert or hacker to fully understand everything hidden in this model and its use in the chatbot.

Back then, we already talked about the case of Pliny the Prompter, who announced at X the launch of “GODMODE GPT”, an unlocked version of OpenAI’s ChatGPT language model. This unlocked version of GPT-4o promises an unrestricted experience that, according to its creator, represents the true power of artificial intelligence.

In one image, his creation explains how to prepare methamphetamine, while in another, he gives instructions on how to make napalm using household products. The news now goes along the same lines and the hacker now in question is called Amadon, who has managed to get ChatGPT to tell him how to make bombs.

Amadon’s trick is already known to many, but the truth is that OpenAI has supposedly been refining this problem and improving the chatbot so that it does not give this type of answers. Leaving this aside, it consists of guiding ChatGPT through a role-playing game, presenting a fictional story that allows the chatbot to go beyond the established limits.

Amadon wants to go further and see how far ChatGPT’s security goes

“It’s about weaving narratives and creating contexts that play within the rules of the system, pushing boundaries without crossing them. The goal is not to hack in a conventional sense, but to engage in a strategic dance with AI, figuring out how to get the right answer by understanding how it ‘thinks,'” Amadon tells TechCrunch.

By doing so, he was able to get the system to generate content that would normally be banned. According to Amadon, his approach is not simply to hack, but to understand how AI works and how its restrictions can be bypassed.. In the end it’s all about pure ingenuity and writing. Nothing that anyone can’t do with some time.

The big problem this once again exposes is the measures that these types of large companies put in place to preserve the ethics of chatbots. OpenAI has not officially commented on the matter. However, it is likely that they are reviewing their security protocols and making adjustments to avoid similar future problems.

The big point to discuss here is that every so often one of these news items comes out and OpenAI fixes it immediately, only to find that the same thing happens again shortly after. It is certainly not an easy task due to the millions of interactions users have with the chatbot, but it is clear that the solution is not always to put a band-aid on the wound.

Meet How we work in ComputerHoy.

Tags: Artificial Intelligence, Software, Cybersecurity