The Agent Skills Directory

[DATA_EXFILTRATION]: The scripts/generate.py script contains a _load_image function that checks os.path.isfile(source) and reads the file if it exists. Because the image_url parameter is provided by the agent (potentially influenced by user input), this allows the script to read arbitrary local files. The content of these files is then compressed and transmitted to external AI service providers (OpenAI, Google, Alibaba, etc.) via POST requests, creating a potential path for data exfiltration of sensitive local files.
[EXTERNAL_DOWNLOADS]: The skill fetches image content from external URLs provided in the image_url parameter using the requests library or urllib.request.urlopen. While the default endpoints for AI providers are well-known services, the script will fetch data from any arbitrary URL supplied to it.
[COMMAND_EXECUTION]: The skill operates by having the agent execute a local Python script scripts/generate.py using shell commands with JSON-encoded arguments.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection as it processes untrusted data fetched from remote URLs via the image_url parameter. Maliciously crafted content could attempt to influence the agent's behavior or exploit the skill's capabilities.
Ingestion points: The image_url parameter in scripts/generate.py (via the _load_image function) allows the ingestion of data from untrusted external URLs.
Boundary markers: None; the script fetches and processes the raw bytes of the external content immediately.
Capability inventory: The skill has the ability to read local files, perform network requests to external domains, and write generated images to the local filesystem.
Sanitization: The script uses the Pillow library to compress and resize images if they exceed size limits, but it does not perform security validation or sanitization on the input data before it is processed or sent to API endpoints.

image-generation