Investigating unintended model actions in our evaluations and internal use

trying to understand the bigger picture here. Anthropic covers the basics but i'd love to hear from people who actually follow this closely