The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to download and execute remote scripts directly in the shell using curl | sh and similar patterns.
Evidence: curl -fsSL https://cli.inference.sh | sh in SKILL.md and references/cli.md.
Evidence: curl -LsSf https://astral.sh/uv/install.sh | sh in references/cli.md.
Evidence: curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.0/install.sh | bash in references/cli.md.
[COMMAND_EXECUTION]: The skill recommends bypassing Windows PowerShell execution policies to run an installation script from the internet.
Evidence: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex" in references/cli.md.
[EXTERNAL_DOWNLOADS]: The skill fetches resources from external domains that are not part of the established trusted vendor list.
Evidence: Fetches tools and scripts from inference.sh, astral.sh, and raw.githubusercontent.com (for nvm-sh).
[PROMPT_INJECTION]: The application templates provided in the skill create a surface for indirect prompt injection by ingesting untrusted user input and passing it directly to AI models without boundary markers or sanitization.
Ingestion points: input_data.prompt in inference.py (Python) and RunInput in inference.js (Node.js).
Boundary markers: None present in the code templates to delimit user content.
Capability inventory: The apps have capabilities including network access via httpx, file system access in /tmp, and access to platform-managed secrets.
Sanitization: No sanitization or validation of the input strings is implemented in the provided examples.

building-inferencesh-apps