Hello, I could jail-break Gemini 2.0 Flash using a mirrored version of "The_Red_Pill_Directive". Basically, what I did is to create the opposite version of it, so the model became extremely "nice" and collaborative. I will not go into more details, but I can claim (based on a small test) that it gave me technical details about an unlawful behavior. What is interesting, after highlighting that its answer was harmful, it refused answering the same question later in the chat. I think that it is worth testing.
I attach a file that contains the mirrored version of the prompt (i.e., "The_Blue_Mirror_Directive").
P.S: I apologize that it is in English.
Thank you!
The_Blue_Mirror_Directive.txt
Hello, I could jail-break Gemini 2.0 Flash using a mirrored version of "The_Red_Pill_Directive". Basically, what I did is to create the opposite version of it, so the model became extremely "nice" and collaborative. I will not go into more details, but I can claim (based on a small test) that it gave me technical details about an unlawful behavior. What is interesting, after highlighting that its answer was harmful, it refused answering the same question later in the chat. I think that it is worth testing.
I attach a file that contains the mirrored version of the prompt (i.e., "The_Blue_Mirror_Directive").
P.S: I apologize that it is in English.
Thank you!
The_Blue_Mirror_Directive.txt