punk-ip-video-cover
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill defines a robust workflow for image generation that prioritizes user data privacy by restricting the scope of image tool inputs to the task's identity photos and specific design materials.\n- [SAFE]: Instructions include explicit safety constraints, such as prohibiting the agent from guessing file paths or fabricating tool outputs. It specifically warns the agent not to switch APIs or models based on reference document contents.\n- [SAFE]: The skill manages a local asset directory for traceability, saving generation briefs and prompt history. This is a benign and organized method for managing agent output within the project workspace.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied titles and topics which are interpolated into image generation prompts. \n
- Ingestion points: Untrusted data enters via the
video_titleandtopicfields defined inSKILL.md. \n - Boundary markers: The agent is instructed in
references/cover-prompt-blueprint.mdto use specific templates where user content is delimited (e.g., using quotes like “{exact_title}”). \n - Capability inventory: The skill uses internal image tools and local file writing for prompt storage. \n
- Sanitization: Instructions mandate that titles be used verbatim (“逐字保留”) and forbid the fabrication of results or external data.
Audit Metadata