The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied delivery items (text, images, metadata), which represents an attack surface for indirect prompt injection.
Ingestion points: Processes external text, visual assets, and metadata as described in SKILL.md and the references/ folder.
Boundary markers: The skill explicitly instructs the agent to use "clear constraints" when using other tools and to "check results line by line" against "real samples."
Capability inventory: The skill identifies and modifies content patterns. No dangerous subprocess calls, dynamic execution, or unauthorized network operations were found.
Sanitization: The instructions emphasize manual review and verification within the final deployment context to ensure no unwanted content or "style drift" is introduced.
[DATA_EXPOSURE]: The skill provides instructions for identifying and removing metadata (EXIF, XMP, C2PA) from delivery items. This is a localized cleanup process for finalized assets intended for publication and does not involve exfiltrating sensitive user information or credentials.
[REMOTE_CODE_EXECUTION]: No patterns of remote code download or execution were found. The skill relies on descriptive instructions for the agent to process content rather than executing external scripts.
[OBFUSCATION]: No hidden text, Base64 encoding, zero-width characters, or homoglyph attacks were detected. The content is provided in plain text and standard YAML configuration.

renwei