<?xml version="1.0" encoding="UTF-8"?><?xml-stylesheet href="/rss.xsl" type="text/xsl"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>fathomforge — writing</title><description>Notes on autonomous agents, runaway cost, and the security of systems that act on their own.</description><link>https://fathomforge.dev/</link><language>en-us</language><copyright>© fathomforge</copyright><managingEditor>carlos@fathomforge.dev (Carlos M.)</managingEditor><atom:link href="https://fathomforge.dev/rss.xml" rel="self" type="application/rss+xml"/><item><title>Three shapes of a runaway</title><link>https://fathomforge.dev/writing/three-shapes-of-a-runaway/</link><guid isPermaLink="true">https://fathomforge.dev/writing/three-shapes-of-a-runaway/</guid><description>Five months of production agent logs, and almost every expensive incident is one of three patterns. None of them look like failures while they&apos;re happening.</description><pubDate>Thu, 10 Sep 2026 00:00:00 GMT</pubDate><category>agents</category><category>cost</category><category>incidents</category></item><item><title>Prompt injection is a blast radius problem</title><link>https://fathomforge.dev/writing/prompt-injection-is-a-blast-radius-problem/</link><guid isPermaLink="true">https://fathomforge.dev/writing/prompt-injection-is-a-blast-radius-problem/</guid><description>You cannot reliably detect the attack. So stop making detection the control, and start bounding what a successful one costs you — including being honest about which bounds you actually have.</description><pubDate>Sun, 06 Sep 2026 00:00:00 GMT</pubDate><category>prompt-injection</category><category>agents</category><category>security</category></item><item><title>Reading usage without reading prompts</title><link>https://fathomforge.dev/writing/reading-usage-without-reading-prompts/</link><guid isPermaLink="true">https://fathomforge.dev/writing/reading-usage-without-reading-prompts/</guid><description>A guardrail sees everything an agent says. Here&apos;s how to build one that enforces a spend cap without ever being able to reconstruct a conversation.</description><pubDate>Wed, 02 Sep 2026 00:00:00 GMT</pubDate><category>privacy</category><category>design</category><category>agents</category></item></channel></rss>