<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Zkml on Lee Madajczyk</title><link>https://madajczyk.com/tags/zkml/</link><description>Recent content in Zkml on Lee Madajczyk</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Fri, 03 Apr 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://madajczyk.com/tags/zkml/index.xml" rel="self" type="application/rss+xml"/><item><title>The Agent Did What?!</title><link>https://madajczyk.com/archive/2026/04/03/the-agent-did-what/</link><pubDate>Fri, 03 Apr 2026 00:00:00 +0000</pubDate><guid>https://madajczyk.com/archive/2026/04/03/the-agent-did-what/</guid><description>&lt;p&gt;&lt;a href="https://madajczyk.com/archive/2026/03/28/familiar-security-failures-ai-acceleration/"&gt;Last week&lt;/a&gt; I wrote about an attack surface that&amp;rsquo;s becoming more and more important: routing and orchestration layers that accumulate credentials and privilege while operating below the security threshold anyone has actually set for them. &lt;a href="https://www.itpro.com/security/litellm-pypi-compromise-everything-we-know-so-far"&gt;LiteLLM&lt;/a&gt; was the clearest example: a routing library holding API keys for every model it touches, compromised through PyPI, &lt;a href="https://www.securityweek.com/mercor-hit-by-litellm-supply-chain-attack/"&gt;4TB exfiltrated&lt;/a&gt; from Mercor. The argument was that these layers are trusted implicitly yet also monitored poorly, and that the security failures compromising them aren&amp;rsquo;t new.&lt;/p&gt;</description></item></channel></rss>