The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates administrative command execution through kubectl exec, oc rsh, and oc debug. Specifically, the oc debug node -- chroot /host command is used to execute operations with host-level privileges on the cluster's underlying nodes. Additionally, the shell scripts reference an external local library ../../../shared/lib/preflight.sh that is not provided for verification.
[DATA_EXFILTRATION]: Provides functionality to retrieve and display cluster administrative credentials, such as rosa show credentials and az aro list-credentials. It also handles sensitive etcd data by creating snapshots and storing them locally on the agent's host filesystem.
[EXTERNAL_DOWNLOADS]: Pulls well-known diagnostic container images, including busybox and nicolaka/netshoot, from public registries for runtime network and DNS troubleshooting. These are neutral references to established technology tools.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted data from cluster resource metadata and logs.
Ingestion points: Data enters the context via kubectl get, oc get, aws logs, and oc logs commands in SKILL.md and scripts/pre-upgrade-check.sh.
Boundary markers: No explicit delimiters or instructions are used to separate untrusted cluster data from the agent's operational logic.
Capability inventory: The agent possesses extensive administrative capabilities, including host-level command execution and the ability to modify cluster-wide configurations.
Sanitization: Log data is processed using grep without sanitization or validation to filter out potential embedded malicious instructions before the data is evaluated by the agent.

cluster-ops