The first version ran ttyd as root with NET_ADMIN, so any player could run nft flush ruleset. The fix moved the privilege into an init container, and the next fix nearly crash-looped every pod over CAP_CHOWN.
Dev and prod shared a cluster. Dev’s cleanup job found prod’s labs missing from its own database, called them orphans, and deleted 28 of them in 30 days. The fix: a missing row is never permission to delete.
A golden-image bake deleted /etc/machine-id. DHCP failed, cloud-init never finished, SSH never opened, and every symptom pointed at netplan. Plus a k3s hostname trap that stranded a pod for 14 hours.