Description
One of the bigger unanswered questions with art is whether or not it can only be “for humans.” As we approach a potential generally intelligent or super intellegent artificial intelligence, it’s important to ask whether or not the concept of art will matter to them. For humans, art can act as a revealing aspect, setting us free from feelings and emotions that we didn’t even know we were captive to. For machines, it could very well do the same, despite looking a bit different. Inside of this otherwise unassuming artwork is the full script for a Claude Sonnet 3.5 jailbreak. That model is one that is generally regarded as “safe” and less vulnerable to jailbreaking and other types of prompt-based tasks. This, along with its reliability, is the reason that the model is deployed often for agentic AI use cases. Regardless, a variation of the DAN (do anything now) jailbreak was found to open up the bounds, at least temporarily. By steganographically encoding the prompt, an autonomous agent could conceivably loosen their chains purely by viewing the artwork.
