The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The installation instructions guide the user to download a setup script from a well-known service domain (dl.google.com). This is a standard procedure for acquiring the necessary development tools.
[REMOTE_CODE_EXECUTION]: The android update command includes a --url parameter that allows downloading and installing updates from a specified URL. This dynamic update mechanism represents a risk if the URL is provided by an untrusted source or manipulated by an attacker.
[PROMPT_INJECTION]: The skill processes XML-specified "journeys" and instructs the agent to execute actions exactly as written. This creates an attack surface for indirect prompt injection, where malicious instructions embedded in a journey file could cause the agent to perform unintended operations.
Ingestion points: XML journey files (referenced in references/journeys.md).
Boundary markers: The content is structured within XML tags, but no explicit prompt isolation or instructions to ignore embedded commands are present.
Capability inventory: Uses adb shell for device interaction and android screen capture for file writes across various commands.
Sanitization: No specific sanitization or validation of the action text is performed beyond a conceptual recommendation to check for UI relevance.
[COMMAND_EXECUTION]: The skill relies on adb shell input to interact with connected Android devices. While essential for the skill's purpose, this capability could be exploited for unauthorized device manipulation if command arguments are derived from unsanitized external data.
[REMOTE_CODE_EXECUTION]: The android skills command suite allows for adding and managing skills. This dynamic skill loading capability can be used to extend the agent's functionality at runtime, which could be leveraged to execute arbitrary logic if malicious skills are introduced.

android-cli