The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (LOW): The skill instructs the user to install puppeteer and playwright. These are well-known packages maintained by trusted organizations (Google and Microsoft respectively), which downgrades the severity per the trust rules.
[COMMAND_EXECUTION] (LOW): Provides instructions for executing shell commands such as npm install and node script.js to facilitate the browser automation environment.
[DATA_EXFILTRATION] (LOW): The skill provides patterns for capturing screenshots (page.screenshot) and extracting page data (page.content(), page.textContent()). While these are intended primary functions, they present a risk surface for data leakage if used on sensitive or untrusted pages.
[PROMPT_INJECTION] (LOW): This skill exposes a surface for Indirect Prompt Injection (Category 8). An attacker-controlled website could include instructions intended to hijack the agent's behavior when the agent processes the page content.
Ingestion points: Untrusted web content enters the agent context through page.goto(), page.content(), and page.textContent().
Boundary markers: Absent. The snippets do not demonstrate use of delimiters to separate web data from agent instructions.
Capability inventory: Full browser automation, network access, and JavaScript execution in the browser context via page.evaluate().
Sanitization: Absent. The skill does not provide methods for sanitizing or validating extracted web content.

chrome