The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Provides instructions to download and execute installer scripts directly from official vendor domains, including the Docker convenience script (get.docker.com), the HuggingFace CLI installer (hf.co), and the AWS CLI installer (awscli.amazonaws.com).
[EXTERNAL_DOWNLOADS]: Includes an optional section to clone and execute a community-developed tool from an unverified third-party GitHub repository (justinwlin/Runpod-Network-Volume-Storage-Tool).
[COMMAND_EXECUTION]: Utilizes sudo for administrative operations such as installing the AWS CLI, updating system packages via apt, and configuring Docker user permissions.
[COMMAND_EXECUTION]: Instructions include using uv run to execute code within the unverified community storage tool after cloning its repository.
[COMMAND_EXECUTION]: Requests Windows administrative privileges to execute PowerShell commands for enabling WSL2.
[CREDENTIALS_UNSAFE]: Guides the setup and storage of sensitive authentication data, including HuggingFace tokens, Docker Hub Personal Access Tokens, S3 API keys, and SSH private keys in standard configuration files (e.g., ~/.aws/credentials, ~/.ssh/id_ed25519).
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection:
Ingestion points: External data is brought into the environment via hf download (HuggingFace Hub) and gh repo clone (GitHub).
Boundary markers: No boundary markers or warnings are present to delineate untrusted external content from instruction data.
Capability inventory: The skill allows for arbitrary command execution via Bash across multiple CLIs, code execution via uv run, and privilege escalation via sudo.
Sanitization: No sanitization or validation of external content (such as repository files or model metadata) is performed before processing.

companion-clis