Anthropic is making auto mode the default in Claude Code, citing research that shows humans catch far fewer dangerous commands than its classifier does.