📄 中文摘要
一种新型攻击类别被识别，利用大型语言模型（LLMs）中上下文推理的核心机制。与传统的提示注入攻击不同，该攻击隐秘地将恶意语言框架嵌入到无害文本中。这些框架在先前上下文中整合时，系统性地改变模型的决策轨迹，而不会触发现有的防御机制。结果是输出的微妙但显著的变化，通过代理管道无声传播，绕过当前的安全措施。该攻击利用了LLMs对语言框架的差异处理，特定框架在上下文中的战略性位置可以有效影响模型的输出。
新攻击类别利用大型语言模型的上下文解读，绕过过滤器：探讨缓解策略
📄 English Summary
New Attack Class Exploits LLM Context Interpretation, Bypassing Filters: Mitigation Strategies Explored
A novel class of attacks has been identified that exploits the core mechanisms of contextual reasoning in large language models (LLMs). Unlike traditional prompt injection attacks that rely on explicit payloads, this class operates covertly by embedding malicious linguistic frames within benign text. When integrated into prior context, these frames systematically alter the model's decision-making trajectory without triggering existing defense mechanisms. The result is a subtle yet significant shift in output, propagating undetected through agentic pipelines and bypassing current security measures. This attack leverages the differential processing of linguistic frames by LLMs, where specific frames, when strategically positioned in context, can effectively influence the model's output.