The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is centered around executing shell commands via the agent-browser CLI. These commands control the user's browser through the Chrome DevTools Protocol, performing actions such as listing tabs, navigating URLs, and interacting with page elements.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it ingests untrusted data from live websites via the snapshot -i and get text body commands. Malicious instructions embedded in web content could theoretically influence the agent's behavior.
Ingestion points: Browser content is retrieved through agent-browser snapshot -i and agent-browser get text body (SKILL.md).
Boundary markers: Explicit boundary markers for external web content are absent in the prompt templates.
Capability inventory: The agent has extensive capabilities including shell command execution (agent-browser), browser-based JavaScript execution (eval), and high-level interaction (click, fill, open).
Sanitization: The skill includes human-in-the-loop requirements, such as 'Confirm with the user before submitting/publishing' and instructions to stop if login or CAPTCHAs are encountered.
[REMOTE_CODE_EXECUTION]: The skill utilizes the eval command of the agent-browser utility to execute JavaScript within the browser's main-frame context. While this is a functional requirement for browser automation and data extraction, it allows for arbitrary code execution in the browser context which should be monitored for misuse.
[EXTERNAL_DOWNLOADS]: The setup instructions recommend installing the agent-browser utility from an official registry. This tool originates from a well-known development organization and is consistent with the skill's stated purpose.

chrome-automation