The Agent Skills Directory

[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection by ingesting untrusted user data and using it to drive agent research capabilities.
Ingestion points: The Intake (M1) step in SKILL.md accepts product names and current prices directly from the user.
Boundary markers: The instructions include a validation step to ensure inputs are present, but do not provide delimiters or instructions for the agent to ignore potentially malicious command-like text within the product name string.
Capability inventory: The skill uses WebSearch and WebFetch tools (detailed in references/sources.md) to retrieve data from external technology media sites.
Sanitization: There are no instructions to sanitize or escape user-provided product names before they are interpolated into the search query templates defined in references/sources.md (e.g., "{商品名} 京东价格 2024 2025").
[DATA_EXFILTRATION]: The skill generates and presents the user with external links to third-party price tracking services such as manmanbuy.com (e.g., https://cu.manmanbuy.com/dingyue/keyword/{关键词}/2). While these links are provided for the user's benefit, this behavior involves transmitting user-supplied search terms to external domains that are not part of the standard whitelisted services.

promo-predictor