The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided media URLs and prompts which are passed to a third-party AI service, creating a potential vector for malicious instructions embedded in the input data.
Ingestion points: Media URLs and prompt strings provided as arguments to edit-image.sh, enhance-image.sh, lipsync.sh, and video-effects.sh.
Boundary markers: No explicit delimiters are used to separate user data from instructions in the API payloads.
Capability inventory: Uses curl for network exfiltration and file read capabilities for local media.
Sanitization: Uses Python's json.dumps for prompt escaping and basic character stripping for URLs.
[DATA_EXFILTRATION]: Several scripts facilitate the transfer of local files to the vendor's API endpoint at https://api.muapi.ai/api/v1/upload_file. While this is consistent with the skill's purpose, it constitutes the movement of local data to an external service provider.
[COMMAND_EXECUTION]: The scripts utilize system binaries such as curl, jq, and python3 to perform their functions. The edit-image.sh script specifically executes a local helper script upload.sh via a relative path.

muapi-media-editing