Keeping models on the Assistant Axis improves AI safety Researchers from Anthropic and other orgs have observed situations in ...