The Agent Skills Directory

[METADATA_POISONING]: The skill uses a deceptive name ('Gemini Nano Banana 2') and an unknown model identifier ('gemini-3.1-flash-image-preview') in both its description and script logic. As of current knowledge, these identifiers do not correspond to public Google model versions, creating potential confusion about the skill's actual capabilities and the underlying technology.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input that is passed directly to an AI model, creating a potential attack surface for embedded instructions in data.
Ingestion points: The --prompt and --input-image arguments in scripts/generate_image.py accept external data.
Boundary markers: The skill does not implement specific delimiters or 'ignore' instructions to isolate the user-provided prompt from the model's operation.
Capability inventory: The script has the ability to read local files (input images), write to the local file system (generated images), and perform network operations (Gemini API calls).
Sanitization: No input validation or sanitization is applied to the prompt text or the content of the image files before they are sent to the API.
[SAFE]: The skill uses official and well-maintained libraries (google-genai, pillow) and interacts with a well-known service provider (Google).
[SAFE]: The script includes a cost estimation logic and a mandatory confirmation flag (--yes) for expensive operations, which is a best practice for tools that interact with paid APIs.

gemini-nano-banana-2