Research question
Do alignment methods make models robust to adversarial inputs, especially multimodal ones?
Key findings
- Text-only attacks were limited by optimisation difficulty, but adversarial images reliably jailbreak multimodal models - Predicts that better text optimisers will break aligned text models