The Agent Skills Directory

[DATA_EXPOSURE]: The script contains logic to search for and load environment variables from sensitive file paths in the user's home directory.
Evidence: The load_env function in scripts/icon_generator.py checks ~/.config/skills/.env and ~/.env. While this is intended for API key management, accessing hidden configuration files in the root of a home directory is a sensitive operation.
[DYNAMIC_EXECUTION]: The script performs dynamic loading of external code by modifying the system path at runtime based on a computed relative path.
Evidence: In scripts/icon_generator.py, the get_background_remove_module function calculates a path to a sibling skill (background-remove/scripts), inserts it into sys.path, and imports the remove_background module. This pattern of loading code from computed paths is a medium-risk execution vector.
[COMMAND_EXECUTION]: The workflow instructions in SKILL.md describe executing shell commands with interpolated user-provided strings, which creates a risk for command injection.
Evidence: The SKILL.md file instructs the agent to execute python3 ... --prompt "your enhanced prompt". If the "enhanced prompt" (derived from user responses to AskUserQuestion) contains shell metacharacters and is not correctly escaped by the agent platform during execution, it could lead to arbitrary command execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input and interpolates it into a structured prompt for a downstream image generation model without explicit sanitization.
Ingestion points: User input is collected via AskUserQuestion for the icon 'Concept' and 'Colors' (SKILL.md).
Boundary markers: No boundary markers or 'ignore' instructions are used when building the icon-specific prompt in Step 2.
Capability inventory: The skill possesses network access (Google API), file system write access (saving icons), and the ability to read local files (reference images).
Sanitization: No sanitization or validation is performed on the user-provided strings before they are formatted into the generation prompt.

icon-generation