Anthropic identifies AI persona drift and ties it to an “assistant axis”; tests across 275 roleplay characters, raising safety limits.

Your team has pulled in data from a variety of sources, integrated it into a shared picture of what’s going wrong, and built a plan of attack. Great start. But now the next challenge begins: How do ...