The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the execution of multiple Node.js scripts (navigate.js, screenshot.js, evaluate.js, etc.) and a shell script (install-deps.sh) to perform its browser automation functions.
[EXTERNAL_DOWNLOADS]: The installation process involves downloading Node.js packages (puppeteer, debug, yargs) from the NPM registry and system libraries via package managers like apt-get on Linux environments to support Chrome's execution.
[REMOTE_CODE_EXECUTION]: The evaluate.js tool allows for the dynamic execution of arbitrary JavaScript code within the browser context via the --script parameter, which can be used to run logic against the DOM of a loaded page.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes content from untrusted external URLs that could contain malicious instructions.
Ingestion points: Data is ingested from arbitrary web pages via scripts like snapshot.js, scrape.js, console.js, and network.js.
Boundary markers: There are no identified boundary markers or explicit instructions to the agent to disregard instructions embedded within the scraped or monitored web content.
Capability inventory: Across the provided scripts, the skill can execute code in the browser context (evaluate.js), write files (screenshot.js), and perform network monitoring (network.js).
Sanitization: The skill documentation does not mention sanitizing, filtering, or escaping the content retrieved from external websites before the agent interprets the resulting JSON output.

chrome-devtools