The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data through the prompt and --input parameters in scripts/generate_image.py.
Ingestion points: The script accepts user-supplied prompts and file paths/URLs through command-line arguments, which are subsequently transmitted to the OpenRouter API.
Boundary markers: The instructions lack explicit delimiters or warnings to the agent regarding potential instructions embedded within prompts or image metadata.
Capability inventory: The script possesses file-read capabilities (path.read_bytes()) and performs network operations (urllib.request.urlopen).
Sanitization: The encode_reference function validates that local files have recognized image extensions (.png, .jpg, .jpeg, .gif, .webp) before processing.
[DATA_EXFILTRATION]: The skill reads local files specified by the agent and transmits their Base64-encoded contents to openrouter.ai. While this is an intended feature for reference-based image generation, it provides a channel for sending local data to an external service.
[EXTERNAL_DOWNLOADS]: The script downloads content from remote URLs provided as inputs or returned by the API response. It implements safety measures including a 64MB size limit and a requirement for HTTPS for URLs originating from the API response.

generate-image