The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions require the agent to ingest and analyze untrusted data from the Kubernetes environment (logs, events, pod descriptions) using commands like kubectl logs and kubectl get events. This creates a surface for indirect prompt injection where malicious content in logs could influence the agent's behavior.
Ingestion points: Diagnostic data is fetched from the cluster via kubectl logs <pod> --previous, kubectl describe pod, and kubectl get events as described in references/kubernetes-debugging.md and references/crash-diagnosis.md.
Boundary markers: There are no instructions to use specific delimiters or protective framing when the agent processes this untrusted external data.
Capability inventory: The agent is authorized to execute complex kubectl commands and provide configuration advice based on the ingested information.
Sanitization: The skill lacks explicit guidance on sanitizing or validating the content of logs and event messages before interpreting them as part of the triage process.
[COMMAND_EXECUTION]: The skill provides instructions for executing highly privileged shell commands intended for deep system troubleshooting.
Evidence: In references/resource-debugging.md, the skill describes using kubectl debug node which creates pods with hostPID, hostNetwork, and root access to the node's filesystem mounted at /host.
Evidence: references/cobalt-kvm-exporter.md details a infrastructure DaemonSet configuration requiring runAsUser: 0, hostPID: true, and SYS_PTRACE capabilities.

kubernetes