The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/install-deps.sh script uses sudo to install system packages on Linux distributions. This requires administrative privileges and modifies the host system's configuration.
[REMOTE_CODE_EXECUTION]: The scripts/evaluate.js script utilizes eval() to execute arbitrary JavaScript code passed as a command-line argument. While this occurs within the browser sandbox, it allows the agent to execute dynamic and potentially untrusted code.
[EXTERNAL_DOWNLOADS]: The skill triggers the download of Node.js dependencies (puppeteer, debug, yargs) and various system libraries through native package managers (apt, dnf, pacman). These downloads originate from official and well-known repositories.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its broad capabilities and data ingestion.
Ingestion points: Data is ingested from external websites via snapshot.js (DOM elements), console.js (logs), and network.js (headers and response bodies).
Boundary markers: The scripts do not implement delimiters or provide instructions to the agent to ignore commands potentially embedded in the ingested web content.
Capability inventory: The skill possesses powerful capabilities including browser control (clicking, form filling) and arbitrary script execution (evaluate.js).
Sanitization: Content retrieved from external websites is returned to the agent in raw JSON format without filtering or sanitizing for malicious instructions.

chrome-devtools