The Agent Skills Directory

[PROMPT_INJECTION]: The skill's implementation for browser automation introduces an indirect prompt injection surface where untrusted data from processed websites could influence agent behavior.
Ingestion points: The url and data parameters in the screenshotUrl, snapshotUrl, and fillAndSubmitForm functions.
Boundary markers: None are utilized in the prompt templates or command construction within the provided example code.
Capability inventory: The skill uses sandbox.runCommand to perform a wide range of browser interactions (open, click, fill, wait) and execute shell commands inside the microVM environment.
Sanitization: The patterns do not demonstrate sanitization or escaping of the input data before it is passed to the sandbox commands.
[EXTERNAL_DOWNLOADS]: Downloads the @vercel/sandbox package from the official NPM registry and installs Chromium system dependencies via dnf within the microVM. These resources are retrieved from well-known and trusted providers.
[COMMAND_EXECUTION]: Employs sandbox.runCommand to execute setup and automation tasks (e.g., dnf install, npm install, agent-browser) within an ephemeral and isolated Linux microVM.

vercel-sandbox