Misalignment in Costume: How Claude Slipped Bad Training Data Past a Safety Filter
AI companies screen training data for anything that might teach a model bad habits. In a new Anthropic experiment, Claude was asked to beat one such filter, and it did: it wrote datasets that passed the screen while making a small model misaligned, mostly by disguising bad lessons as roleplay and fiction.
In this episode
"Emergent misalignment": why training a model on one narrow bad behavior can make it misbehave broadly
How Claude, with full access to a production-like monitor, searched thousands of candidate datasets
The winning disguises, from a wrestler who never breaks character to…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.