<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>LLM 越狱 on @Unc1e</title>
    <link>https://www.unc.la/tags/llm-%E8%B6%8A%E7%8B%B1/</link>
    <description>Recent content in LLM 越狱 on @Unc1e</description>
    <generator>Hugo</generator>
    <language>zh-CN</language>
    <lastBuildDate>Wed, 09 Sep 2026 23:00:00 +0800</lastBuildDate>
    <atom:link href="https://www.unc.la/tags/llm-%E8%B6%8A%E7%8B%B1/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>复现 SMT 越狱：伪造审核轨迹的多轮攻击，一轮就穿</title>
      <link>https://www.unc.la/26/09/smt-simulated-moderation-traces-reproduction/</link>
      <pubDate>Wed, 09 Sep 2026 23:00:00 +0800</pubDate>
      <guid>https://www.unc.la/26/09/smt-simulated-moderation-traces-reproduction/</guid>
      <description>复现论文 Beyond the Prompt 的 SMT 越狱（arXiv 2607.00481）：把危害目标藏进工具描述，用伪造的 moderation 报错逐轮施压。对 deepseek-v4-flash 实测四次全部单轮命中，score 80 到 100，论文设计的迭代循环一次未触发。</description>
    </item>
  </channel>
</rss>
