The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to use powerful administrative tools including kubectl exec and kubectl debug. These commands allow the agent to obtain interactive shell access to running containers and inject diagnostic workloads, which represents a high-privilege capability in the cluster environment.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the processing of cluster data.
Ingestion points: The agent is instructed to ingest potentially attacker-controlled data via kubectl logs, kubectl describe pod (event logs), and kubectl get pods -A (field selectors).
Boundary markers: The instructions lack boundary markers or warnings to the agent to disregard natural language instructions that might be embedded within container logs or event messages.
Capability inventory: The agent possesses terminal access with the ability to execute arbitrary commands inside containers (kubectl exec) and apply resource manifests (kubectl apply), which could be abused if the agent is manipulated by malicious log content.
Sanitization: There is no requirement for the agent to sanitize or filter the output of log commands before interpreting the content as diagnostic information.

kubernetes-specialist