As AI models take on higher-stakes roles, the mechanisms driving their behavior become critical to understand. We found that emotion vectors are implicated in some of Claude’s most concerning failure modes.

→ View original post on X — @anthropicai