The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements tools for executing arbitrary commands through the kubectl binary, specifically via the raw and exec functions in skill.js. The raw tool allows for the construction of unconstrained command strings, while the exec tool permits running commands directly within containerized environments. These features provide the agent with significant administrative control over the cluster.
[CREDENTIALS_UNSAFE]: The config tool provides an action=view parameter that retrieves and displays the content of the active kubeconfig. This configuration file contains sensitive information, including cluster API endpoints, client certificates, and authentication tokens required for cluster access.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by processing untrusted data from the cluster environment.
Ingestion points: The skill retrieves pod logs and resource descriptions (metadata and specifications) from the cluster via the logs, get, and describe tools defined in skill.js.
Boundary markers: Data returned from cluster operations is not enclosed in protective delimiters or accompanied by instructions for the agent to ignore embedded commands.
Capability inventory: The skill possesses capabilities to modify cluster state using the apply, create, and delete tools, and to execute code in pods via the exec tool.
Sanitization: Content fetched from the cluster, such as log output or resource specifications, is passed to the agent without validation or escaping, potentially allowing malicious content within those fields to influence agent behavior.

kubernetes