The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection by ingesting and processing data from potentially untrusted sources within a Kubernetes cluster.
Ingestion points: The agent ingests data from cluster events, pod logs, and performance metrics via commands like az aks agent --query as described in SKILL.md and references/examples.md.
Boundary markers: There are no documented boundary markers or instructions designed to separate ingested data from agent instructions or to treat cluster data as untrusted text.
Capability inventory: As documented across multiple files (e.g., references/api-server-etcd.md and references/kubelet-logs.md), the agent has capabilities to execute kubectl commands, access raw logs, and suggest or potentially execute remediation actions like deleting webhooks.
Sanitization: The documentation does not indicate that data retrieved from the cluster is sanitized or validated before being processed by the underlying LLM.
[REMOTE_CODE_EXECUTION]: Documents the installation of CLI tools using remote script execution from a trusted source.
Evidence: The skill provides instructions for installing the Azure CLI using curl -sL https://aka.ms/InstallAzureCLIDeb | sudo bash in references/troubleshooting.md.
[EXTERNAL_DOWNLOADS]: Fetches configuration files from a trusted repository to enable monitoring features.
Evidence: The skill uses wget to download a Prometheus collector configuration map from Cloudflare's official GitHub organization in references/control-plane-metrics.md.
[COMMAND_EXECUTION]: The skill documents extensive use of CLI tools including az and kubectl to manage and troubleshoot cluster resources.
Evidence: Numerous commands for cluster diagnostics, log retrieval, and configuration modification are detailed throughout all reference files.

az-aks-agent