3 ms·
Google Bard: Image to Prompt Injection
- elromulous 3y agoThis seems consistent with previous attacks (albeit with a different modality). I think this cements the attacks, and shows they can be done irrespective of modality.
- cjbprime 3y agoIt'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.