The Agent Skills Directory

[SAFE]: The skill and its associated script perform standard administrative tasks such as labeling namespaces and observing network flows using established CLI tools (kubectl and hubble). The instructions emphasize GitOps practices for configuration changes, which is a security best practice.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection attack surface due to its interaction with live cluster data.
Ingestion points: The scripts/hubble-debug.sh script and manual commands in SKILL.md ingest output from hubble observe and kubectl get, which include metadata like pod names and labels that could be influenced by a cluster-level attacker.
Boundary markers: The instructions lack explicit delimiting or warnings to the agent regarding the untrusted nature of the data retrieved from the cluster environment.
Capability inventory: The agent is authorized to read cluster state (hubble observe, kubectl get) and modify it (kubectl label for the emergency escape hatch).
Sanitization: While jq is used in the helper script to structure the output, there is no semantic validation of the content being returned to the agent's context.

network-policy