The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions describe constructing shell commands (e.g., helm template <release> <chart-path>) using variables directly derived from user-provided $ARGUMENTS. If the input is not properly sanitized or escaped by the agent platform, it could allow for shell command injection through the use of metacharacters like semicolons or backticks.
[PROMPT_INJECTION]: The skill ingests and renders external data from Kubernetes manifests and cluster state, which constitutes an indirect prompt injection surface. Malicious instructions embedded in manifest metadata, comments, or data fields could attempt to manipulate the agent's summary or risk assessment.
Ingestion points: Reads local manifest files (YAML), Helm charts, Kustomize configurations, and live cluster data via kubectl get.
Boundary markers: None. There are no instructions provided to the agent to treat the rendered manifest content as untrusted or to ignore instructions contained within it.
Capability inventory: Uses the Bash tool to execute system binaries including kubectl, helm, diff, mktemp, and cat.
Sanitization: Includes a 'Secret Masking' step that redacts values associated with keys like password, secret, and token before display. However, it does not validate the content for instructions or command patterns.

k8s-diff