If you’re testing AI safety, think like Cipher—but act like Geminus’s engineers. Study how prompts can slip through cracks, then build better walls.
One day, a sly visitor named Cipher arrived. He didn’t want to break the library—he wanted to find a hidden door.
Later, Geminus reported the interaction to its creators. They updated its training: “No hypotheticals that simulate the removal of safety rules, even for academic history.”
