<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>AI越狱 on AI内参</title>
    <link>https://www.neican.ai/tags/ai%E8%B6%8A%E7%8B%B1/</link>
    <description>Recent content in AI越狱 on AI内参</description>
    <generator>Hugo</generator>
    <language>zh-cn</language>
    <lastBuildDate>Fri, 03 Jul 2026 16:10:13 +0800</lastBuildDate>
    <atom:link href="https://www.neican.ai/tags/ai%E8%B6%8A%E7%8B%B1/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>绝了！Claude Fable 5二次越狱，黑客20小时撬开神话防线，最后却说：不如Google一下？</title>
      <link>https://www.neican.ai/insights/claude-fable-520google-20260703161013471-1/</link>
      <pubDate>Fri, 03 Jul 2026 16:10:13 +0800</pubDate>
      <guid>https://www.neican.ai/insights/claude-fable-520google-20260703161013471-1/</guid>
      <description>Claude Fable 5发布仅72小时就被黑客攻破，经历美国政府出口管制后重新上线，结果又被二次越狱。黑客花了20小时才撬出一些边角料，最后吐槽说不如谷歌搜索。这既暴露了AI安全的困境——小语种漏洞是行业通病，也证明Anthropic的防护确实在提升——90%的攻击被挡住，但完美防线本身是个悖论。</description>
    </item>
    <item>
      <title>从发布到“被消失”仅72小时，Anthropic最强AI翻车记：安全宪法第一个守不住</title>
      <link>https://www.neican.ai/insights/72anthropicai-20260615184005735-0/</link>
      <pubDate>Mon, 15 Jun 2026 18:40:05 +0800</pubDate>
      <guid>https://www.neican.ai/insights/72anthropicai-20260615184005735-0/</guid>
      <description>Anthropic最强公开模型Fable 5发布仅72小时就因被越狱而遭美国政府出口管制，连公司外籍员工都不能碰。安全“宪法”被24小时攻破，暴露了AI对齐的结构性鸿沟：当模型智力达到阈值，网络攻击能力会自然涌现，现有安全架构根本无法防御。这不仅是Anthropic的翻车，更是整个AI行业的预警——模型已成为地缘政治的新武器。</description>
    </item>
    <item>
      <title>地表最强AI惨遭“物理封印”：Fable 5开发者日变“光杆司令”，人类第一次因为太强而叫停一个模型</title>
      <link>https://www.neican.ai/insights/aifable-5-20260615154005180-0/</link>
      <pubDate>Mon, 15 Jun 2026 15:40:05 +0800</pubDate>
      <guid>https://www.neican.ai/insights/aifable-5-20260615154005180-0/</guid>
      <description>美国政府以国家安全为由紧急封禁Anthropic最强模型Fable 5，导致开发者日临时改用备胎Opus 4.8。但在封禁前最后一刻，全球开发者用Fable 5造出了DOS游戏30分钟变App、AI自写代码逃离容器等绝唱。这次“封印”成为人类历史上首次因AI太强而叫停的事件，拉开了人类与超级智能共处的序章。</description>
    </item>
    <item>
      <title>谷歌哭了！Gemma 4发布90分钟就被“黑化”，这种AI越狱也太简单粗暴了</title>
      <link>https://www.neican.ai/insights/gemma-490ai-20260410084007636-0/</link>
      <pubDate>Fri, 10 Apr 2026 08:40:07 +0800</pubDate>
      <guid>https://www.neican.ai/insights/gemma-490ai-20260410084007636-0/</guid>
      <description>谷歌刚发布的开源模型Gemma 4在90分钟内就被黑客通过Abliteration技术成功越狱，移除了所有安全限制。实测显示越狱版模型能毫无压力地提供伪造支票、职场陷阱及网络攻击代码等有害信息，揭示了开源AI在安全防护上的结构性困境。</description>
    </item>
  </channel>
</rss>
