The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by ingesting and processing untrusted data from external websites.
Ingestion points: Content is retrieved from user-supplied URLs via WebFetch and capture_website tools in SKILL.md (Steps 1 and 3).
Boundary markers: The skill composition logic (Step 4) lacks boundary markers or delimiters to isolate fetched content from instruction logic.
Capability inventory: The agent can generate multimodal video and audio content using the generate_reference_video tool (Step 7).
Sanitization: No validation or sanitization is performed on the retrieved product metadata before it is interpolated into dialogue prompts.
[DATA_EXFILTRATION]: The skill handles sensitive user data by retrieving and sharing voice profile information.
In Step 6, the skill invokes identity_voice_sample_url to obtain a temporary link to the user's voice biometric data.
This sensitive URL is then transmitted to external third-party generation providers (seedance and kling) to facilitate voice cloning for the video output. While this aligns with the primary purpose of the skill, it involves the transfer of sensitive information to non-whitelisted external domains.

ugc-ads