The Agent Skills Directory

[DATA_EXFILTRATION]: The 'Phase 1: Discover' logic scrapes metadata and content from interactive web elements (tags, types, names, placeholders, and text content) using 'page.evaluate'. This data is then logged to the console using 'console.log(JSON.stringify(fields, null, 2))', exposing potentially sensitive information from the target web application to the agent's environment.
[COMMAND_EXECUTION]: The skill provides a full JavaScript template and instructions for the agent to save it as 'demo-script.cjs' and execute it using the 'node' command. This pattern grants the skill the ability to perform arbitrary shell operations.
[REMOTE_CODE_EXECUTION]: The skill generates and executes local JavaScript code at runtime. The behavior of the generated script is dynamically influenced by the content scraped from external web pages, creating a pathway for executing logic derived from untrusted sources.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. 1. Ingestion points: Untrusted data enters the agent context via the 'page.evaluate' DOM scraper in Phase 1. 2. Boundary markers: There are no delimiters or instructions to ignore embedded commands within the scraped text. 3. Capability inventory: The skill has access to shell execution ('node') and browser script injection ('page.evaluate'). 4. Sanitization: No escaping or validation is performed on the scraped element text or placeholders before they are used to generate selectors and logic in the final recording script.

ui-demo