The Agent Skills Directory

[Data Exposure & Exfiltration] (HIGH): The file references/deployment.md contains instructions for using a specific private SSH key path: ~/.ssh/linux-ai. Referencing or accessing sensitive identity files is a high-risk activity.
[Credential Safety] (HIGH): The file references/env-and-keys.md details the management of several sensitive credentials, including LITELLM_MASTER_KEY, EXA_API_KEY, and VLLM_STUDIO_API_KEY. While described as setup steps, providing instructions for secret management within an agent's context increases the risk of accidental exposure.
[Command Execution] (MEDIUM): The skill documentation (references/deployment.md) instructs the user or agent to execute ./start.sh and perform process management (pkill). Executing unverified local scripts allows for arbitrary code execution on the host machine.
[Indirect Prompt Injection] (LOW): The skill defines a surface for processing external data through the /chat UI and tool-call prompts.
Ingestion points: /chat UI (mentioned in references/deployment.md).
Boundary markers: Absent; there are no instructions to the agent to treat external inputs as untrusted.
Capability inventory: Shell script execution (./start.sh), SSH connection (ssh), and Docker service management.
Sanitization: Absent; no sanitization or validation of the input used in tool calls is described.

vllm-studio