The Agent Skills Directory

Indirect Prompt Injection (MEDIUM): The skill processes untrusted external data (prompts and image URLs) to generate media outputs, creating an attack surface where malicious inputs could influence agent behavior or downstream processes.
Ingestion points: The prompt, image_url, and end_image_url arguments in tools such as submit_task, sync_generate_image, and upload_image (SKILL.md).
Boundary markers: None identified in the provided instructions; untrusted data is passed directly into tool arguments.
Capability inventory: The skill facilitates network communication with external AI platforms (e.g., fal.ai, Jimeng) to generate and return media files.
Sanitization: No sanitization or validation logic for the provided URLs or prompt text is present in the skill definition.
Data Exposure & Exfiltration (LOW): The upload_image tool accepts an image_url parameter. If the underlying MCP server lacks strict validation, this could be used for Server-Side Request Forgery (SSRF) to probe internal network resources.

sutui-ai