The Agent Skills Directory

[DATA_EXFILTRATION]: The skill hardcodes absolute file paths to a specific user's local directory and iCloud synchronization folder: /Users/zz/Library/Mobile Documents/iCloud~md~obsidian/Documents/zhangAI/Clippings/Reading/. This exposes the internal system username and directory structure, and indicates the tool is tightly coupled to a specific environment.
[REMOTE_CODE_EXECUTION]: The script scripts/collect_web_pack.py utilizes dynamic code execution to load a secondary script at runtime. It uses importlib.util to locate collect_web_research_pack.py based on computed paths and executes its contents into the current process memory via exec_module. While targeting local files, this pattern of dynamic execution increases risk if file resolution or path integrity is compromised.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core function of scraping untrusted external web content and converting it into Markdown.
Ingestion points: Untrusted data enters the environment from arbitrary URLs via requests.get calls in scripts/collect_web_research_pack.py.
Boundary markers: The skill uses YAML frontmatter (e.g., source, role) to separate metadata from the scraped body in the generated output files, although external content is interpolated directly into the Markdown body.
Capability inventory: The skill has the capability to perform network requests (via requests and the Jina AI reader service) and write files to the local filesystem (Markdown files and an assets/ directory).
Sanitization: The skill implements sanitization using BeautifulSoup and readability-lxml to strip dangerous elements such as scripts, navigation blocks, and advertisements before final processing.

z-web-pack