The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image data through external models, returning extracted text and layout information to the agent context. This establishes a surface for indirect prompt injection if images contain text specifically crafted to override agent instructions. * Ingestion points: Model results from Moondream and Florence-2 are returned to the agent via scripts/moondream.py and referenced in the skill's instructions. * Boundary markers: There are no explicit delimiters surrounding the ingested model output in the provided prompt templates. * Capability inventory: The skill is capable of file system operations (read/write) and network requests. * Sanitization: Text returned from external models is not sanitized before being presented to the agent.
[DATA_EXFILTRATION]: The utility script scripts/moondream.py reads from the user's home directory environment file (~/.env) to locate the FAL_KEY. While functional for locating the service credential, accessing sensitive files in the home directory is a documented data exposure risk.
[EXTERNAL_DOWNLOADS]: The skill downloads generated image masks from the fal.ai CDN (fal.media) when using the segmentation features. This operation uses the urllib.request module to fetch resources from the service's official infrastructure. The script includes a specific security audit suppression comment (noqa: S310) for this network operation.

fal-ai