LLMs Open to Manipulation Using Doctored Images, Audio

As LLMs begin to integrate multimodal capabilities, attackers could use hidden instructions in images and audio to get a chatbot to respond the way they want, say researchers at Black Hat Europe 2023.