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

If an attacker splits a malicious prompt into discrete chunks, some large language models (LLMs) will get lost in the details and miss the true intent.