The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script scripts/heiyan-booklist-scraper.js programmatically extracts session authentication tokens from browser cookies.
Evidence: The getToken function in scripts/heiyan-booklist-scraper.js retrieves the Admin-Token using document.cookie.match(/Admin-Token=([^;]+)/). While the skill documentation states this is for calling the target platform's API, the ability to extract session tokens is a significant credential exposure risk.
[COMMAND_EXECUTION]: The skill executes external binary commands on the host operating system to facilitate web scraping.
Evidence: The ab function in scripts/cdp-utils.js uses child_process.execFileSync to invoke the agent-browser binary with various arguments. The skill also performs file system write operations via fs.writeFileSync to save reports.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external web platforms, creating an indirect prompt injection surface where malicious instructions in story descriptions or titles could influence the agent's behavior.
Ingestion points: Web content scraped from ishugui.com and manage.zhangwenpindu.cn via the scraper scripts.
Boundary markers: Data is structured in Markdown format but lacks explicit delimiters or specific instructions to the agent to ignore embedded commands within the scraped text.
Capability inventory: The agent performs market analysis and summary tasks based on the ingested data, which are then used to influence writing decisions.
Sanitization: No sanitization or filtering is performed on text fields such as book titles or descriptions before they are presented to the agent.
[DYNAMIC_EXECUTION]: The skill constructs and executes JavaScript code dynamically within a browser context.
Evidence: The evalJSONBase64 function in scripts/cdp-utils.js sends Base64-encoded JavaScript strings to the agent-browser for evaluation in the target web page context.

story-short-scan