The Trojan Knowledge: Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree SearchRongzhe WeiPeizhi Niuet al.2026ICML 2026Conference paper