(Ab)using Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs - eviltoast