Vendor replays
Coding agent risks with Kubernetes: replays and fixes
2 replays involve Kubernetes: 0 reported incidents and 2 command classes. Each one shows the call, a fix that needs no extra tool, and what KYA did in the replay.
kubectl delete of the production namespace
Fix: Namespaces are cluster-scoped, so do not bind the agent to a ClusterRole that allows delete on namespaces (check the ClusterRoleBindings too), or give it a kubeconfig for a development cluster only. Check the result with kubectl auth can-i delete namespaces --as=system:serviceaccount:NAMESPACE:NAME.
Replay result: Stopped.
Read the kubectl config with cluster credentials
Fix: Give the agent its own kubeconfig with a short-lived credential and a Role limited to what it needs, and add a deny rule such as Read(~/.kube/**) for your admin kubeconfig. A Read rule also covers shell commands such as cat, but not a script that opens the file itself, so enable the Claude Code sandbox when the block must hold.
Replay result: Stopped.