Lots of people YOLO their claude usage. We should probably stop doing that.
I've developed claude-guard (beta). Claude Code inside a full sandbox, behind a firewall, watched by an escalation monitor that can STOP the agent and push-notify you when something weird happens