Twitter user Pliny the Prompter, a self-described ethical hacker, released a custom version of OpenAI’s flagship neural network called GODMODE GPT last Wednesday. He managed to make the new GPT-4o model ignore all restrictions: it used profanity, explained how to break into cars and how to prepare banned substances.

Of course, the experiment could not continue for long. When the GODMODE GPT neural network gained relative viral popularity on social network X, the incident was brought to the attention of OpenAI and the custom model was removed from the site just hours after its launch. It is no longer possible to access it, but screenshots containing “malicious hints” of GPT-4o were preserved in the author’s original branch in the X social network.

It is stated that the GPT-4o model may have been attacked using old internet jargon leetspeak (replacing letters with numbers and special characters when typing), as confirmed by screenshots. OpenAI did not respond when asked if this jargon could be used to bypass ChatGPT restrictions. It is possible that the GODMODE GPT that created it liked leetspeak and hacked the system in another way.

The incident has become a manifestation of the large-scale AI red teaming movement, in which ethical hackers identify weaknesses in modern AI systems before they can cause significant harm to them. They have incredible capabilities, but as recent experiences with Google and search reviews show, today’s generative AI is still a system that makes a good guess about which words should appear next in a text, without having any real intelligence.