The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to retrieve sensitive information using commands like kubectl get secret [NAME] -n wtr-supplier-charges -o yaml. This action exposes the contents of Kubernetes secrets, including potentially sensitive configuration and credentials, to the agent's processing context.
[COMMAND_EXECUTION]: The workflows utilize the Bash tool to execute high-impact commands such as kubectl patch, kubectl scale, and gcloud projects add-iam-policy-binding. These commands allow the agent to modify the cluster state and GCP project IAM policies.
[COMMAND_EXECUTION]: Instructions suggest reading from local files, specifically $(cat key.json), which typically contains sensitive service account keys or credentials, for use in creating Kubernetes registry secrets.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted data from Kubernetes logs and events.
Ingestion points: kubectl logs, kubectl describe, and kubectl get events commands in SKILL.md and examples/examples.md.
Boundary markers: Absent. The instructions do not define delimiters or provide warnings to the agent regarding the content of logs.
Capability inventory: Extensive capabilities via Bash, including kubectl (read/write/patch) and gcloud (resource and IAM management).
Sanitization: Absent. Logs and descriptions are displayed and analyzed without filtering.

gcp-gke-troubleshooting