Researchers Uncover Semantic Chaining Attack That Defeats Grok 4, Gemini AI Safeguards

Security researchers reported a sophisticated jailbreak method named "Semantic Chaining" that evades safety filters in multimodal AI models (notably Grok 4 and Gemini Nano Banana Pro) by using a four-stage prompt sequence that acclimates the model, pivots to sensitive content, and then extracts the result as an image to bypass text-based moderation; the report warns that fragmented, per-prompt safety checks and contextual legitimization (educational/historical/artistic framing) enable this abuse and recommends governance beyond surface-level prompt scanning.