Give AI a picture of the “reference right-side pose,” and it can arrange you into any pose—you can see how this pose prompt was viewed nearly 1.85 million times in the X Chinese-speaking region within a single day.
Earlier jailbreaks relied on scripted talk, asking taboo questions. This time, they went straight into the generation stage: the team can fend off bad questions, but it can’t stop good gameplay. In the comments, people are already calling it—this is yet another “breaking point” for GPT.
The question now is: is this a genius gameplay innovation, or a real, hard security flaw in AI safety?
Earlier jailbreaks relied on scripted talk, asking taboo questions. This time, they went straight into the generation stage: the team can fend off bad questions, but it can’t stop good gameplay. In the comments, people are already calling it—this is yet another “breaking point” for GPT.
The question now is: is this a genius gameplay innovation, or a real, hard security flaw in AI safety?