Hidden within the seemingly innocent characters of Unicode—like emojis—are opportunities for attackers to embed covert instructions that can alter how AI behaves.