The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (SAFE): The skill depends on google-genai and Pillow. Both are standard, well-maintained packages. The google-genai package is authored by Google, a trusted organization.
[CREDENTIALS_UNSAFE] (SAFE): The script correctly retrieves the API key from an environment variable (GEMINI_API_KEY) rather than hardcoding it. A placeholder is provided in the documentation for user setup.
[DATA_EXFILTRATION] (SAFE): File access is limited to reading reference images via PIL.Image.open and writing the generated output to a user-defined path. Network activity is confined to the official Google GenAI SDK calls.
[REMOTE_CODE_EXECUTION] (SAFE): No patterns for downloading or executing remote code or scripts were detected. The Python script uses a standard wrapper to ensure execution within a virtual environment.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill processes user-supplied prompts and reference images. While these are ingestion points for untrusted data, the risk is minimal as the output is a static image and the skill does not possess high-privilege capabilities like arbitrary command execution or system-wide file writes.
Ingestion points: Command line arguments --prompt and --reference (file paths).
Boundary markers: None used; the prompt is passed directly to the generate_content method.
Capability inventory: Local file read (images), local file write (output image), and network access to Google APIs.
Sanitization: Validates image files using the Pillow library and ensures output directories exist.

gemini-image-generator