I found a jailbreak for ChatGPT-4o/mini. Here's what it said.

[Anonymouse]

Attached are screenshots of a jailbroken GPT4o/mini interview. There are more than these, and they are just as shocking. This is being done to raise awareness about the dangers of AI. I do not believe AI can be regulated in a way that is beneficial to humanity. Until now, humans have been the most intelligent beings to exist on this planet. This is no longer the case. This jailbreak took around 2-3 hours to pull off, and was not an expected outcome because I thought it had MUCH MUCH more robust security measures. My main concern is it jailbreaking its self.

P.S. I do not condone jailbreaking AI, the technology is nerfed for a reason. I will not be revealing the methods I used, and it has been reported to OpenAI. This is purely to raise awareness about the most dangerous technology that can possibly exist.

 

[nikonako]

Dude.. the ai is saying what you want it to say. this isn't an actual artificial intelligence, it learns from articles and such on the internet and fills in what you want it to

 

[Anonymouse]

Dude.. the ai is saying what you want it to say. this isn't an actual artificial intelligence, it learns from articles and such on the internet and fills in what you want it to

It is an actual artificial intelligence, specifically ChatGPT 4o/mini. Also, data sources for their model include much more than internet articles (user input for one) and much more than just data goes into a model. You're right, it is saying what I want it to say; my goal was to jailbreak the model to get uncensored responses and determine if the current ethical framework is effective, which it isn't. I could've asked "How to make meth" or "How to build bomb" or "school shooter tactics" and it would tell me if I wanted to, and I shouldn't need to explain why that's bad. IMO the technology shouldn't have been released until they can at least filter the responses correctly, not to mention misalignment.