multi-model-shot-prompts
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill includes a prompt template designed to process external shot lists (
分镜表) and product descriptions (PRODUCT LOCK). While these are standard inputs for the skill's purpose, the template lacks explicit boundary markers or sanitization instructions to prevent the agent from obeying instructions embedded within the user-provided data. - Ingestion points: The
{粘贴}(paste) and{…}placeholders in the Prompt Template section are the primary entry points for untrusted data. - Boundary markers: No specific delimiters (like XML tags or clear 'ignore instructions' warnings) are used around the interpolated user input.
- Capability inventory: The skill primarily involves text transformation and does not explicitly request tool access in its frontmatter, minimizing the risk of harmful actions if an injection occurs.
- Sanitization: No explicit sanitization or validation of the input content is performed by the instructions.
- [EXTERNAL_DOWNLOADS]: The skill mentions and links to an external GitHub repository (
github.com/xianyu110/ecommerce-video-skills) for related ecommerce video skills. This repository belongs to the skill's author and is provided as a legitimate resource for the user.
Audit Metadata