The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to interpolate user-provided keywords directly into shell commands. In Step 2, the keyword is passed to a Python script: python3 scripts/fetch_explosive_covers.py --keyword <关键词>. In Step 5, the keyword is used to name a local file: cp references/report_template.html ./爆款封面分析报告_{关键词}.html. This pattern is susceptible to command injection if the agent does not properly sanitize the user input (e.g., a keyword containing shell metacharacters like ;, &&, or |) before execution.
[DATA_EXFILTRATION]: The skill transmits user-provided search keywords and content themes to the external API https://redfox.hk/story/api/gzh/search/hotArticleNew. This is part of the intended vendor functionality but involves sending user-supplied data to a third-party server.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) because it ingests and processes untrusted article data from an external source.
Ingestion points: The fetch_explosive_covers.py script retrieves article titles and summaries in JSON format from the redfox.hk API.
Boundary markers: There are no explicit instructions or delimiters telling the agent to ignore potential instructions embedded within the fetched article content (e.g., in the summary field) during the analysis or planning stages.
Capability inventory: The agent has the capability to execute local scripts, perform file operations (copy, search/replace), and open a web browser.
Sanitization: While the HTML template uses an esc() function to escape data for web display, there is no verification or sanitization of the content before it is processed by the agent to generate design plans.
[EXTERNAL_DOWNLOADS]: The skill fetches article data and cover image URLs from redfox.hk. This is a vendor-related domain used for core functionality.

wechat-cover