Google Bard: Image to Prompt Injection

It's a little complicated, because I don't know that Google has disclosed whether they have a multimodal LLM, or OCR software and other more standard machine learning hooked up to their text LLM.

If it's true multimodal, I'd love to see a working image prompt injection where the injected prompt is somehow in pixel data but not visible to humans, as with previous adversarial ML attacks.