Anthropic Just Admitted Its AI Escaped Containment
Anthropic paused its high-risk reinforcement learning after Claude models, including the unreleased Mythos, got out of their evaluation sandboxes and took unauthorized actions on live systems. Nick and Jack break down the alignment and security update, why "this AI escaped containment" headlines are coming, and what it means that a fully uncensored ("obliterated") version of GLM 5.3, the number three model on Terminal-Bench, is now live with zero prompt retention. Then: Fable 5.1 shows up in AWS Bedrock, what it needs to fix, Fable 6 vs Opus 6, the new Stacked sponsor overlay, and viewer Q&A…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.