LLMs Open to Manipulation Using Doctored Images, Audio : vim

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.

Related Keywords

Vitaly Shmatikov , Ben Nassi , Google Deepmind , Harry Potter , Eugene Bagdasaryan , Cornell University , Helmholtz Center , Sequire Technology , Information Security At Saarland University , Black Hat Europe , Indirect Instruction Injection , Black Hat , Tsung Yin Hsieh , Information Security , Saarland University , Harry Potter Like ,