Claude Opus 4.6: System Card Part 1: Mundane Alignment and Model Welfare

The Don’t Worry About the Vase Podcast is a listener-supported podcast. To receive new posts and support the cost of creation, consider becoming a free or paid subscriber.

* 00:00:00 - Introduction

* 00:03:50 - A Three Act Play

* 00:04:53 - Safety Not Guaranteed

* 00:10:53 - Pliny Can Still Jailbreak Everything

* 00:13:01 - Transparency Is Good: The 212-Page System Card

* 00:14:07 - Mostly Harmless

* 00:18:47 - Mostly Honest

* 00:21:10 - Agentic Safety

* 00:23:29 - Prompt Injection

* 00:27:38 - Key Alignment Findings

* 00:38:10 - Behavioral Evidence (6.2)

* 00:43:33 - Reward Hacking and ‘Overly Agentic Actions’

* 00:45:53 - Metrics (six point two point five point two)

* 00:48:01 - All I Did It All For The GUI

* 00:49:14 - Case Studies and Targeted Evaluations Of Behaviors (6.3)

* 00:49:36 - Misrepresenting Tool Results

* 00:50:29 - Unexpected Language Switching

* 00:52:34 - The Ghost of Jones Foods

* 00:54:15 - Loss of Style Points

* 00:55:07 - White Box Model Diffing

* 00:55:26 - Model Welfare

https://open.substack.com/pub/thezvi/p/claude-opus-46-system-card-part-1?utm_campaign=post-expanded-share&utm_medium=web