The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructs the agent to handle Hugging Face API tokens by creating a Kubernetes secret using a shell command with the token passed as a literal string.
Evidence: The command kubectl create secret generic hf-token --from-literal=HF_TOKEN="<user-provided-token>" -n <namespace> is provided in Step 1 of the deployment process.
Risk: Passing secrets via command-line arguments can expose them in shell history, process listings, and system logs.
[PRIVILEGE_ESCALATION]: The provided Kubernetes deployment template explicitly allows the container to run with root privileges.
Evidence: The securityContext in templates/vllm-deployment.yaml specifies runAsNonRoot: false.
Risk: Running containers as root increases the potential impact of a container breakout or vulnerability exploitation.
[COMMAND_EXECUTION]: The skill operates primarily by executing shell commands via kubectl to manage cluster resources.
Evidence: Multiple steps in SKILL.md involve running kubectl apply, kubectl get, and kubectl delete.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from the user (such as tokens, model names, and namespaces) and interpolates them directly into shell commands.
Ingestion points: User-provided strings for the Hugging Face token, Kubernetes namespace, and model name during the deployment and customization steps.
Boundary markers: Absent. The instructions use simple placeholders without explicit sanitization or delimiters for the shell commands.
Capability inventory: The skill uses kubectl to create, modify, and delete resources in the Kubernetes cluster.
Sanitization: Absent. There is no logic to validate or escape inputs before they are used in the command line.
[EXTERNAL_DOWNLOADS]: The skill pulls a container image from the vendor's repository.
Evidence: The deployment uses the vllm/vllm-openai:latest image.
Context: This is the standard distribution method for the vLLM project.

vllm-deploy-k8s