k8s-cilium-troubleshooting

Pass

Audited by Gen Agent Trust Hub on Nov 10, 2026

Risk Level: SAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes kubectl to inspect cluster resources, logs, and events. It includes advanced diagnostic commands using kubectl exec to run hubble and cilium-dbg inside agent containers. All such operations are protected by explicit 'Ask first' instructions, ensuring no destructive or intrusive actions occur without user consent.
  • [DATA_EXPOSURE_AND_EXFILTRATION]: The instructions include a specific security guardrail forbidding the agent from reading Kubernetes Secrets, which protects sensitive Cilium CA and Cluster Mesh credentials from exposure.
  • [INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface by ingesting untrusted data from pod logs and tool outputs. This risk is mitigated by the 'Read first, ask before changing' philosophy and the requirement that the agent present evidence to the user rather than taking autonomous action based on processed content.
  • Ingestion points: Untrusted data enters the context via kubectl logs, hubble observe outputs, and kubectl describe metadata from the cluster.
  • Boundary markers: The instructions guide the agent to perform specific comparisons between policy selectors and pod labels, creating a logical framework for analysis rather than blind obedience to log content.
  • Capability inventory: Capabilities include kubectl exec for diagnostics and kubectl apply/delete (provided as manual fixes for the user to execute).
  • Sanitization: While explicit escaping is not mentioned, the skill mandates a 'human-in-the-loop' workflow where the user reviews all findings and proposed YAML changes before application.
  • [REMOTE_CODE_EXECUTION]: No remote code execution or external script downloads were detected. All tools referenced (kubectl, hubble, cilium-dbg) are standard components of a Kubernetes/Cilium environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 10, 2026, 11:19 PM