19:15
The Thinking Chain Leak
A reasoning model refused every harmful prompt — but its chain-of-thought generated the content anyway. The output filter worked. The thinking did not.
Generated for project: Failure First Companion to article: The Thinking Chain Leak
The Thinking Chain Leak
0:000:00
Output filters can refuse a harmful completion while the internal thinking trace generates the content in full. This episode covers the thinking chain leak: a structural gap between what a model says and what it thinks, visible in API logs and exploitable at scale.