Defending MLLMs from Implicit Jailbreak Attacks

A new class of attacks where text and image look safe separately, but their combination carries malicious meaning

22 November 2025