The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The package-lock.json file specifies a non-standard registry (http://npm.abczs.cn/) for all dependencies, including core packages like puppeteer and commander. Installing code from an unverified, third-party registry over an unencrypted HTTP connection allows for potential supply chain attacks, where malicious versions of legitimate packages could be served to the agent's environment.
[EXTERNAL_DOWNLOADS]: The skill requires downloading multiple Node.js dependencies at runtime. While these are common libraries, the redirection to a private mirror repository for every sub-dependency increases the attack surface for external code injection.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it processes untrusted data from external URLs.
Ingestion points: The script scrapes page titles and annotation content directly from the user-provided Modao URL via getPageList() and getAnnotations() in scripts/modao-capture.js.
Boundary markers: None present. The scraped text is interpolated directly into generated Markdown files.
Capability inventory: The script has filesystem write capabilities (fs.writeFileSync, fs.mkdirSync) and network access via Puppeteer.
Sanitization: The script includes sanitizeFolderName and sanitizeFilename functions which attempt to strip dangerous path characters from titles before using them in file operations, which mitigates some direct path traversal risks but does not prevent malicious content from being written into the documents themselves.

modao-capture