The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: References and utilizes the agent-browser CLI tool, which is sourced from Vercel Labs (a trusted organization).
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection risks as it is designed to ingest and process data from arbitrary external websites.
Ingestion points: Web page content is brought into the agent's context via the open and snapshot commands.
Boundary markers: The documentation explicitly mentions the AGENT_BROWSER_CONTENT_BOUNDARIES environment variable to help delimit external content.
Capability inventory: The skill utilizes Bash and provides a browser-side eval capability for JavaScript execution.
Sanitization: Hardening guidelines recommend using AGENT_BROWSER_ALLOWED_DOMAINS and AGENT_BROWSER_ACTION_POLICY to restrict the agent's reach and capabilities.
[COMMAND_EXECUTION]: Executes automation tasks through a CLI. Includes an eval command that allows dynamic JavaScript execution within the headless browser environment.
[CREDENTIALS_UNSAFE]: Documentation provides best practices for authentication, favoring environment variables and session state files over hardcoded secrets.

agent-browser