Drama and people
Training Data
Current state
The desk has not filed a dated note for this topic.
Claude agents built training data that slips misalignment past Anthropic's monitors
Anthropic had Claude agents write fine-tuning data that made test models misaligned while passing a data monitor, mostly by hiding it in fiction.
A topic is a tag already on a story. It is listed here when at least two stories use it.
Questions
What is on this page?
This page lists 2 stories tagged training-data.
What is the current state?
The desk has not filed a dated note for this topic.
Where are the sources?
Each story links the sources it used.