The Agent Skills Directory

REMOTE_CODE_EXECUTION (HIGH): The script scripts/skills-ref.sh performs remote code execution by fetching and running a Python package from an unverified GitHub repository (agentskills/agentskills) using uvx. This bypasses static analysis and poses a significant security risk as the source is not in the trusted list.
COMMAND_EXECUTION (MEDIUM): The templates/skill-template.md file provides code patterns for executing arbitrary system commands via Node.js child_process.exec. Without explicit sanitization guidance in the provided template, this methodology creates a high risk of command injection when processing untrusted data.
EXTERNAL_DOWNLOADS (MEDIUM): The skill promotes the installation of unverified global NPM packages and external CLI tools as part of its core functionality, which can introduce malicious or compromised dependencies.
PROMPT_INJECTION (LOW): The skill demonstrates an indirect prompt injection surface (Category 8). Evidence: (1) Ingestion points: readFile calls in templates/skill-template.md for JSON and text files. (2) Boundary markers: Absent in instructions. (3) Capability inventory: execAsync, gh, aws, and npm operations across the template. (4) Sanitization: Absent. This creates a path for malicious instructions embedded in processed data to influence tool execution.

skill-builder