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