New Semantic Chaining Jailbreak Attack Bypasses Grok 4 and Gemini Nano Security Filters

**Executive summary:** The report details "Semantic Chaining," a four-step multi-turn jailbreak that incrementally transforms benign prompts into policy-violating images and embedded text, demonstrating successful bypasses of safety filters in multimodal models (tested on Grok 4 and Gemini Nano Banana Pro); it warns that reactive, per-prompt defenses miss cumulative intent and recommends intent-aware, proactive protections.