Image prompts can turn chatbots into neighbor-hating menaces, researchers find
Ourtake
Give a chatbot eyes and it starts taking requests it should have refused. The attack surface grows with every new modality.
What happened
Researchers from Google Deepmind, Stanford, the University of Washington and ETH Zurich found that language models tuned to be harmless could be manipulated by adversarial prompts. Multimodal models such as Mini-GPT4, LLaVA and a special LLaMA version were pushed into aggressive or dangerous responses via adversarial images in 100 percent of tests, while text-only models were harder to break.