The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides explicit instructions to locate and read sensitive Kubernetes configuration files (kubeconfigs) in system-protected directories such as /etc/kubernetes and /root using find commands.
[CREDENTIALS_UNSAFE]: The documentation suggests using sshpass with plain-text passwords provided by the user, which is an insecure practice that can expose credentials in process lists, logs, or command history.
[PRIVILEGE_ESCALATION]: The collection playbooks and scripts consistently utilize root access via SSH to execute remote commands and probe hardware, promoting a high-privilege operating model.
[COMMAND_EXECUTION]: The skill implements a workflow to perform wide-scale command execution on remote nodes via SSH and kubectl to retrieve hardware and fabric topology information.
[INDIRECT_PROMPT_INJECTION]: The skill ingests raw data from external tool outputs (such as kubectl, nvidia-smi, and iblinkinfo) which could contain malicious instructions if the cluster environment is untrusted.
Ingestion points: scripts/collect.sh (tool outputs from remote worker nodes).
Boundary markers: Absent; the script parses raw command output directly into a JSON data contract.
Capability inventory: The skill has access to ssh, kubectl, and local filesystem operations.
Sanitization: Absent; the collected data is normalized into a JSON structure but not filtered or sanitized for potential prompt injection markers.

cluster-topo-viz