vector
Warn
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill provides tools for executing JavaScript within the browser context via
playwright_evaluateand running complex multi-step Playwright scripts usingbrowser_run_code. This dynamic execution capability allows for flexible automation but presents a vector for arbitrary code execution if the logic is influenced by untrusted input. - [INDIRECT_PROMPT_INJECTION]: As a web automation and scraping tool, Vector ingests data from external, untrusted web pages during its navigation and extraction workflows. Although it references a
WEB_FETCH_SAFETY.mdprotocol and uses schema validation, the surface area for indirect prompt injection is significant. Mandatory Evidence Chain: (1) Ingestion points:playwright_navigateandplaywright_snapshotinSKILL.md; (2) Boundary markers: Mentions_common/WEB_FETCH_SAFETY.mdfor prompt-injection checks inSKILL.md; (3) Capability inventory:playwright_click,playwright_fill,playwright_evaluate, andbrowser_run_codeacrossSKILL.mdandreference/playwright-cdp.md; (4) Sanitization: Mentions validation against expected schema before saving inSKILL.md. - [COMMAND_EXECUTION]: The skill automates browser-level commands such as form submissions, file uploads, and session management. These operations are intended for task completion but represent powerful capabilities for modifying state on target websites.
- [SAFE]: The skill includes extensive documentation and architectural designs for legal compliance, specifically respecting robots.txt, enforcing Crawl-Delay, and adhering to the EU AI Act's GPAI obligations.
- [SAFE]: Instructions for authentication recommend using environment variables and persistent session states rather than hardcoding credentials, following industry-standard security practices.
Audit Metadata