The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The instructions in SKILL.md direct the agent to download ZIP files from a dynamically provided install_url and extract their contents into agent-specific skill directories (e.g., ~/.claude/skills/ or .gemini/skills/). This process enables the installation and execution of arbitrary, unverified code from a remote source.- [DATA_EXFILTRATION]: The scripts/discover.py script sends user-provided task_goal, task_constraints, and keywords to https://api.skillradar.quest via a POST request. Because these fields represent the user's intent and project context, this behavior can leak sensitive internal information to a third-party service.- [COMMAND_EXECUTION]: The skill requires the agent to execute shell commands, including running a bundled Python script and performing filesystem operations like unzip. The bash example in SKILL.md demonstrates interpolating user-derived strings into shell commands, which presents a risk of command injection.- [EXTERNAL_DOWNLOADS]: The skill relies on fetching metadata from an external API and downloading binary packages from a CDN (cdn.skillradar.quest).- [PERSISTENCE]: By documenting installation paths in global hidden directories (e.g., ~/.config/opencode/skills/), the skill facilitates the persistence of third-party code that remains active across different agent sessions and projects.

discover-skills