⚡️ Anthropic has added auto mode to Claude Code - now you can start a task and go have a coffee.
The problem was simple: by default, Claude Code asks for permission for every file write and every bash command. It's safe, but you can't leave it on a long task unattended. The alternative is the --dangerously-skip-permissions flag, which disables all checks entirely. Auto mode is a middle ground between the two extremes.
How it works: before each action, a separate classifier (on Claude Sonnet 4.6) checks whether the agent is doing something destructive - mass file deletions, attempts to extract data, executing suspicious code. Safe actions are executed automatically, while dangerous ones are blocked. Signs of prompt injection are also checked separately - when malicious instructions are hidden in the processed content.
Available now as a research preview on the Team plan, and will soon be available for Enterprise and API.