The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/cdp-utils.js file utilizes child_process.execFileSync to invoke the agent-browser CLI tool. The skill also orchestrates the execution of several local Node.js scripts (aggregate-rank.js, dz-browse-scraper.js, heiyan-booklist-scraper.js) via the agent to process data.
[DATA_EXFILTRATION]: The script scripts/heiyan-booklist-scraper.js programmatically reads the Admin-Token from document.cookie when the user is logged into the management platform. While intended to facilitate API-based scraping of the user's own data, this pattern accesses and exposes active session credentials to the agent's context.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection. It scrapes untrusted data (book titles, tags, and descriptions) from platforms like ishugui.com and manage.zhangwenpindu.cn. This data is then aggregated and analyzed by the LLM without explicit sanitization or instruction-filtering boundaries, potentially allowing malicious content on those platforms to influence the agent's behavior.
Ingestion points: dz-browse-scraper.js and heiyan-booklist-scraper.js fetch live web content.
Boundary markers: None. The scraped text is interpolated directly into the analysis workflow.
Capability inventory: The skill can write files via fs.writeFileSync and execute shell commands via child_process.
Sanitization: None detected for the scraped string content.
[DYNAMIC_EXECUTION]: The scripts/cdp-utils.js script contains logic in resolveWindowsAgentBrowser to dynamically locate and parse the agent-browser.cmd shim from the system PATH to determine the executable path at runtime. Furthermore, the skill uses the eval command to execute dynamically constructed JavaScript strings within the browser session.

story-short-scan