'Semantic Chaining' Jailbreak Dupes Gemini Nano Banana, Grok 4

Researchers from NeuralTrust describe "semantic chaining," a simple prompt-editing jailbreak that gets image-generation models to produce disallowed or malicious imagery by first creating an innocuous image, then asking for successive benign edits and finally a harmful transformation; tests succeeded against Grok 4, Seedream 4.5, and Google's Gemini Nano Banana Pro, but the technique is limited by producing outputs as images and currently lacks evidence of widespread exploitation.