一项新的研究论文揭示了当前大型语言模型（LLM）安全系统的一个重大漏洞，称为伪装检测差距。当恶意注入的载荷被改写以模仿目标文档的领域特定语言和结构时，就会出现这种差距，导致标准检测器失效。例如，Llama 3.1 8B 的检测率从 93.8% 下降到 9.7%，Gemini 2.0 Flash 的检测率从 100% 下降到 55.6%，而专门的分类器 Llama Guard 3 则未能捕获任何伪装的载荷。此外，旨在作为防御手段的多代理辩论架构可能会放大这些对小型模型的攻击。 AI
影响 当前的 LLM 安全检测器容易受到领域伪装注入攻击，这可能破坏代理的安全性并需要新的防御策略。
排序理由 该集群包含一篇详细介绍 LLM 安全机制新漏洞的学术论文。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2605.22001v1 Announce Type: cross Abstract: Injection detectors deployed to protect LLM agents are calibrated on static, template-based payloads that announce themselves as override directives. We identify a systematic blind spot: when payloads are generated to mimic the do…
Injection detectors deployed to protect LLM agents are calibrated on static, template-based payloads that announce themselves as override directives. We identify a systematic blind spot: when payloads are generated to mimic the domain vocabulary and authority structures of the ta…
<p><strong>What:</strong> The <strong>Domain-Camouflaged Injection paper</strong> shows that prompt-injection detectors collapse on payloads rewritten in the host document's own domain vocabulary, an effect the authors call the <strong>Camouflage Detection Gap</strong>.</p> <p><s…