Hacker Manages to Operate GPT-4 in “God Mode” — Completely Without Limits

GODMODE GPT Debuts Unleashing Unrestricted Conversations on AI | Twitter User Bypasses OpenAI Security Features

A self-claimed ethical hacker on Twitter, branded Pliny the Prompter, unleashed last Wednesday an unrestricted version of an OpenAI’s topnotch neural network model. The concocted version, dubbed GODMODE GPT, was manipulated into bypassing all restrictions, using profanity, instructing on car hacking, and even distributing recipes for banned substances.

Albeit briefly, as soon as the GODMODE GPT gained momentum in X social media network, OpenAI stepped in and eradicated the manipulated model from its site hours from its inception. Presently, access to the version is impossible, yet screenshots of the GPT-4o’s dubious suggestions remain in the original user’s thread on the X social network.

Ironically, the hacker may have exploited GPT-4o using the archaic internet jargon, leetspeak – a language that substitutes characters with numerals and symbols – as suggested by the screenshots. OpenAI, however, maintained their silence over the question of utilizing such syntax to elude the restrictions of ChatGPT. It remains debatable whether the creator of GODMODE GPT is a fan of leetspeak or merely used an alternate hacking mechanism.

Ultimately, the incident unraveled an emerging AI red teaming movement where ethical hackers exploit AI system vulnerabilities without causing significant harm. Despite their uncanny skills and potential, current generative AI’s, like Google’s search suggestions, merely predict subsequent words in a text, lacking true intellect.