<?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>Ablation on DL4Sec</title>
    <link>https://dl4sec.com/tags/ablation/</link>
    <description>Recent content in Ablation on DL4Sec</description>
    <generator>Hugo -- 0.147.7</generator>
    <language>en-us</language>
    <lastBuildDate>Fri, 07 Aug 2026 21:27:38 +0200</lastBuildDate>
    <atom:link href="https://dl4sec.com/tags/ablation/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Hack the Bus: Removing Refusal from an Open-Weight LLM</title>
      <link>https://dl4sec.com/blog/generative-ai/hack-the-bus/</link>
      <pubDate>Fri, 07 Aug 2026 21:27:38 +0200</pubDate>
      <guid>https://dl4sec.com/blog/generative-ai/hack-the-bus/</guid>
      <description>Post-training safety looks robust until you find the one direction in activation space that carries refusal. Ablate it and the model stops saying no, which tells us alignment is thinner than it appears.</description>
    </item>
  </channel>
</rss>
